TTS/STT/STS | Friends of the Crustacean 🦞🤝 | Page 1

keen venture Apr 24, 2026, 4:17 AM

#

I want to set a voice for my Bot.

Discord call
Reply back in a voice message on Telegram/Discord. I will send her voice message and she will reply back in voice.

I'll set this up using Eleven Labs and OpenAI

I need a step-by-step, one-step-at-a-time process.

OpenClaw 2026.4.22 (00bd2cf)
WSL2 Win 11

charred gladeBOT Apr 24, 2026, 4:17 AM

#

unreal socketBOT Apr 24, 2026, 4:17 AM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

charred gladeBOT Apr 24, 2026, 4:17 AM

#

spark impBOT Apr 24, 2026, 4:20 AM

#

I’m checking the exact docs paths so I don’t give you made-up keys.

#

You’re setting up 2 different features:

Discord voice channel / call = channels.discord.voice
Telegram/Discord voice-note replies = tools.media.audio + messages.tts

Also, do not install @openclaw/voice-call for this. That plugin is for phone calls (Twilio/Telnyx/Plivo), not Discord voice channels.

Step 1 only: install audio tooling on your WSL2 gateway host.

sudo apt update
sudo apt install -y ffmpeg
ffmpeg -version | head -1
ffprobe -version | head -1

Why first: Discord voice messages need ffmpeg and ffprobe on the gateway host.

#

Docs:

Reply with done or paste the error, then I’ll give you Step 2, the exact OpenAI + ElevenLabs config.

keen venture Apr 24, 2026, 4:21 AM

#

Done

spark impBOT Apr 24, 2026, 4:22 AM

#

Nice. Step 2: put your API keys in ~/.openclaw/.env on the WSL machine running the Gateway.

mkdir -p ~/.openclaw
nano ~/.openclaw/.env

Paste this:

OPENAI_API_KEY=your_openai_key_here
ELEVENLABS_API_KEY=your_elevenlabs_key_here

Notes:

ELEVENLABS_API_KEY can also be XI_API_KEY, but I’d use ELEVENLABS_API_KEY.
No quotes, one key per line.

#

Why this way: if OpenClaw runs as a service, shell env vars often do not carry over, but ~/.openclaw/.env does.

Docs:

https://docs.openclaw.ai/help/faq#env-vars-and-env-loading

Reply done and I’ll give you Step 3, the exact config block.

keen venture Apr 24, 2026, 4:24 AM

#

Hold up. I'm new. Put a dummy key and also show super step by step and also API setting on Eleven Labs, like which option to turn on. ANd also I can choose any voices?

spark impBOT Apr 24, 2026, 4:26 AM

#

Let me sanity-check the ElevenLabs side so I don’t send you on a UI scavenger hunt.

#

Yep, let’s slow it way down.

Answer first: yes, you can use any ElevenLabs voice that exists in your account/library and that your API key can access.
For your first setup, I strongly recommend a premade voice from ElevenLabs, not a cloned/custom one. Less friction.

Step 2, beginner version

Goal for this step only: collect 3 things

OpenAI API key
ElevenLabs API key
One ElevenLabs Voice ID

Do not edit OpenClaw config yet.

2A. Get your OpenAI API key

#

Open: https://platform.openai.com/api-keys
Click Create new secret key
Give it a name like: openclaw-tts
Copy it somewhere safe

It will look roughly like this:

OPENAI_API_KEY=sk-proj-REPLACE_THIS_WITH_YOUR_REAL_KEY

2B. Get your ElevenLabs API key

Log in to ElevenLabs
Go to the dashboard/account area
Open API Keys

#

Click Create API Key
Name it something like: openclaw-voice
Copy the key

It will look roughly like this:

ELEVENLABS_API_KEY=sk_1234567890_REPLACE_ME

What to turn on in ElevenLabs?

For basic OpenClaw TTS, you usually only need:

an API key
a voice

You do not need to set up:

#

Agents
Conversational AI
Webhooks
Phone
STT inside ElevenLabs

If ElevenLabs asks about scopes/restrictions, basic text-to-speech access is what matters.

2C. Pick a voice in ElevenLabs

Open Voices / My Voices / Voice Library
Pick a premade voice you like
Find the three dots or More actions
Click Copy voice ID

It will look something like:

#

ELEVENLABS_VOICE_ID=EXAVITQu4vr4xnSDxMaL

That sample ID is just an example, not “the” required voice.

Dummy example only

Do not use these exact values:

OPENAI_API_KEY=sk-proj-REPLACE_THIS_WITH_YOUR_REAL_KEY
ELEVENLABS_API_KEY=sk_1234567890_REPLACE_ME
ELEVENLABS_VOICE_ID=EXAVITQu4vr4xnSDxMaL

#

Quick answer to “can I choose any voices?”

Yes, if:

the voice is in your ElevenLabs account/library
your key can access it

For first setup, choose one premade voice.

Docs:

Reply with done once you have those 3 values, and I’ll give you Step 3: exactly where to paste them in WSL with a safe beginner config.

keen venture Apr 24, 2026, 4:32 AM

#

done collecting.

The voice is jqcCZkN6Knx8BJ5TBdYR

spark impBOT Apr 24, 2026, 4:33 AM

#

Perfect. The voice ID is not secret, so jqcCZkN6Knx8BJ5TBdYR is fine to use later.
Do not paste your API keys here.

Step 3

Put your 2 API keys into the OpenClaw env file on WSL.

3.1 Open WSL terminal

3.2 Run this

mkdir -p ~/.openclaw
nano ~/.openclaw/.env

3.3 Paste this into the file

#

OPENAI_API_KEY=sk-proj-PASTE_YOUR_REAL_OPENAI_KEY_HERE
ELEVENLABS_API_KEY=sk_XXXX_PASTE_YOUR_REAL_ELEVENLABS_KEY_HERE

3.4 Save the file

In nano:

Press Ctrl+O
Press Enter
Press Ctrl+X

3.5 Lock down the file

chmod 600 ~/.openclaw/.env

That’s it for this step.

#

Docs: https://docs.openclaw.ai/help/faq#env-vars-and-env-loading

Reply done, and I’ll give you Step 4: the exact openclaw.json config for:

Discord voice channel voice
Telegram/Discord voice-note replies
using your ElevenLabs voice ID jqcCZkN6Knx8BJ5TBdYR

keen venture Apr 24, 2026, 4:37 AM

#

done

#

Done

#

@spark imp

spark impBOT Apr 24, 2026, 4:52 AM

#

I’m here 🦐

Step 4

Goal: write the basic voice config, without touching your API keys again.

Paste these in WSL, one by one:

openclaw config set messages.tts.auto "inbound"
openclaw config set messages.tts.provider "elevenlabs"

openclaw config set messages.tts.providers '{
  elevenlabs: {
    voiceId: "jqcCZkN6Knx8BJ5TBdYR",
    modelId: "eleven_multilingual_v2"
  },

#

  openai: {
    model: "gpt-4o-mini-tts",
    voice: "alloy"
  }
}' --strict-json --merge

openclaw config set tools.media.audio.enabled true --strict-json
openclaw config set tools.media.audio.models '[
  {
    provider: "openai",
    model: "gpt-4o-transcribe"
  }
]' --strict-json

openclaw config set channels.discord.commands.native true --strict-json

#

openclaw config set channels.discord.voice.enabled true --strict-json
openclaw config set channels.discord.voice.tts.provider "elevenlabs"

openclaw config set channels.discord.voice.tts.providers '{
  elevenlabs: {
    voiceId: "jqcCZkN6Knx8BJ5TBdYR",
    modelId: "eleven_multilingual_v2"
  },
  openai: {
    model: "gpt-4o-mini-tts",
    voice: "alloy"
  }
}' --strict-json --merge

openclaw config validate

#

What this does

Telegram / voice-note replies: reply in audio only after you send audio
Discord voice channel / call: enables /vc join
Speech-to-text: uses OpenAI gpt-4o-transcribe
Voice: uses your ElevenLabs voice jqcCZkN6Knx8BJ5TBdYR

Important note

This is the clean baseline.

Telegram should do native voice-note style nicely.
Discord voice channels should speak with this voice.
Discord text-chat audio replies may first show up as an audio attachment instead of the fancy waveform-style voice message. We can handle that after the base test.

Docs:

#

Reply with either:

“validate passed”
or paste the exact error output

Then I’ll give you Step 5.

keen venture Apr 24, 2026, 4:57 AM

#

done somehow, got some error too

spark impBOT Apr 24, 2026, 4:58 AM

#

No worries, you’re close.

Step 5 only: run this in WSL and paste the full output:

openclaw config validate

If it says valid, reply: validate passed
If it shows errors, paste the whole error block
Don’t paste API keys, but this command normally won’t show them

Why this step: we need to separate config syntax problems from actual voice/runtime problems before testing.

keen venture Apr 24, 2026, 4:58 AM

#

openclaw config validate                                                                                                                                                                                                                        🦞 OpenClaw 2026.4.22 (00bd2cf) — Your .zshrc wishes it could do what I do.                                                                                                                                                                                             Config valid: ~/.openclaw/openclaw.json

spark impBOT Apr 24, 2026, 4:59 AM

#

Nice, that’s the big hurdle. 🎉 Your config is valid.

Step 6

Restart the gateway so it picks up the new ~/.openclaw/.env API keys:

openclaw gateway restart

Then paste the output, or if it’s clean just reply:

restart done

Why this step: config hot-reloads, but new env keys usually need a gateway restart.

keen venture Apr 24, 2026, 4:59 AM

#

restart done

spark impBOT Apr 24, 2026, 5:00 AM

#

Nice. Step 7: test a one-off TTS reply in Discord.

In this thread, try:

/voice audio Hello from OpenClaw

If /voice doesn’t show up yet, use the text command instead:

/tts audio Hello from OpenClaw

Expected result

You should get back an audio reply.

#

Reply with just one of these:

worked
or paste the exact error

Docs: https://docs.openclaw.ai/tools/tts

keen venture Apr 24, 2026, 5:02 AM

#

Talk voice is not configured.

Missing: talk.provider and talk.providers.<provider>.
Set it on the gateway, then retry.

spark impBOT Apr 24, 2026, 5:02 AM

#

That error is actually helpful, we hit the wrong voice feature. Let me verify the command mapping so I don’t send you in circles.

#

Ah, good catch, that means Discord routed you into Talk mode, not reply TTS.

So for now, ignore /voice.

Step 7, retry

Use this exact text command in chat:

/tts audio Hello from OpenClaw

Important

Do not use /voice for this test
/voice hit the Talk feature
We’re testing message TTS replies first

Expected result

#

You should get back an audio reply using your configured voice.

Reply with one of:

worked
or paste the exact error

If this works, I’ll do the Discord call/voice channel step next.

keen venture Apr 24, 2026, 5:03 AM

#

Same

#

/tts shows the /voice

spark impBOT Apr 24, 2026, 5:04 AM

#

Got it. Then the Discord command routing is probably the problem, not the TTS config itself.

I gave you the wrong test surface earlier, sorry.

Step 8

Run this in WSL, then paste the full output:

openclaw infer tts convert --text "hello from openclaw" --output /tmp/hello.mp3 --json

What this does

tests TTS directly on the gateway
bypasses Discord slash/text command weirdness
tells us whether ElevenLabs/OpenAI + your config are actually working

Reply with:

#

the full JSON output, or
convert worked

Docs: https://docs.openclaw.ai/cli/infer#tts

keen venture Apr 24, 2026, 5:05 AM

#

openclaw infer tts convert --text "hello from openclaw" --output /tmp/hello.mp3 --json                      11:05:16 [plugins] 1 plugin(s) failed to initialize (validation: xai). Run 'openclaw plugins list' for details.                     {                                                                                                                                     "ok": true,                                                                                                                         "capability": "tts.convert",                                                                                                        "transport": "local",                                                                                                               "provider": "openai",                                                                                                               "attempts": [                                                                                                                         {                                                                                                                                     "provider": "elevenlabs",                                                                                                           "outcome": "failed",                                                                                                                "reasonCode": "provider_error",                                                                                                     "latencyMs": 2167,                                                                                                                  "error": "elevenlabs: ElevenLabs API error (401): This request exceeds your API key (Miko) quota of 10. You have 0 credits remaining, while 5 credits are required for this request. [code=quota_exceeded]"                                                           },                                                                                                                                  {                                                                                                                                     "provider": "openai",                                                                                                               "outcome": "success",                                                                                                               "reasonCode": "success",                                                                                                            "latencyMs": 4403                                                                                                                 }                                                                                                                                 ],                                                                                                                                  "outputs": [                                                                                                                          {                                                                                                                                     "path": "/tmp/hello.mp3",                                                                                                           "format": "mp3",                                                                                                                    "voiceCompatible": false

spark impBOT Apr 24, 2026, 5:06 AM

#

Nice, this tells us a lot.

What happened

Your TTS setup works
ElevenLabs failed because your account has no credits
OpenClaw fell back to OpenAI successfully

So the problem is not config anymore. It’s just this:

ElevenLabs ... quota_exceeded ... 0 credits remaining

The xai warning is unrelated for this task, we can ignore it for now.

Next step, choose 1

I recommend Option A for now so you can keep moving.

Option A, recommended

#

Use OpenAI as primary for now, since it already works.

Option B

Add ElevenLabs credits, then keep ElevenLabs as primary.

Reply with:

A = switch primary voice to OpenAI now
B = keep ElevenLabs, I’ll give you the next test after you add credits

If this solved part of it, please right-click the solution message → Apps → ✅ Mark Solution so Answer Overflow can index it.

keen venture Apr 24, 2026, 5:07 AM

#

I got balance, Total
23,736 credits
Remaining
23,642

spark impBOT Apr 24, 2026, 5:07 AM

#

Perfect, then this is not an account-balance issue.

The error is saying your specific API key named “Miko” has its own quota cap of 10 credits, and that key is out, even though your account still has credits.

Step 9

In ElevenLabs → API Keys:

find the key named Miko
either raise/remove its credit limit
or create a new API key just for OpenClaw with a bigger quota / no tiny cap

Do not paste the key here.

When that’s done, reply done, and I’ll give you the next single step to update WSL and retest.

Docs:

https://docs.openclaw.ai/tools/tts

keen venture Apr 24, 2026, 5:08 AM

#

Oh yeah i set 10 credit limit on the API key. I removed it now. What to do now?

spark impBOT Apr 24, 2026, 5:08 AM

#

Nice, that should do it.

Step 10

Run the same WSL test again:

openclaw infer tts convert --text "hello from openclaw" --output /tmp/hello.mp3 --json

What we want to see

Ideally this time the JSON shows:

"provider": "elevenlabs"

or at least an attempts entry where ElevenLabs is success.

#

Paste the output again, or just say:

elevenlabs worked
or paste the error if it still fails

No gateway restart should be needed here since you changed the quota on the provider side.

keen venture Apr 24, 2026, 5:09 AM

#

openclaw infer tts convert --text "hello from openclaw" --output /tmp/hello.mp3 --json                      11:09:03 [plugins] 1 plugin(s) failed to initialize (validation: xai). Run 'openclaw plugins list' for details.                     {                                                                                                                                     "ok": true,                                                                                                                         "capability": "tts.convert",                                                                                                        "transport": "local",                                                                                                               "provider": "elevenlabs",                                                                                                           "attempts": [                                                                                                                         {                                                                                                                                     "provider": "elevenlabs",                                                                                                           "outcome": "success",                                                                                                               "reasonCode": "success",                                                                                                            "latencyMs": 2321                                                                                                                 }                                                                                                                                 ],                                                                                                                                  "outputs": [                                                                                                                          {                                                                                                                                     "path": "/tmp/hello.mp3",                                                                                                           "format": "mp3_44100_128",                                                                                                          "voiceCompatible": false                                                                                                          }                                                                                                                                 ]                                                                                                                                 }

spark impBOT Apr 24, 2026, 5:10 AM

#

Nice, ElevenLabs is working end to end now. 🎉

Next, pick one:

1. Test voice replies
Send the bot a real voice message (not a slash command).

Telegram: should reply as a native voice note
Discord text chat: will likely reply as audio, but not always the fancy waveform voice-message UI

2. Test Discord voice channel / call
We’ll do the /vc join step next.

I recommend 1 first.

Reply with 1 or 2.

keen venture Apr 24, 2026, 5:15 AM

#

Telegram is working. No reply on Discord.

Also, the voice is so different the one I selected

spark impBOT Apr 24, 2026, 5:16 AM

#

Good, Telegram working means the core TTS pipeline is healthy.

The different voice part is very likely because Discord can use a separate override from normal reply TTS:

normal reply TTS: messages.tts
Discord voice playback: channels.discord.voice.tts

Step 11

Run this one command in WSL and paste the output:

printf 'messages.tts\n'; openclaw config get messages.tts --json; printf '\n\nchannels.discord.voice.tts\n'; openclaw config get channels.discord.voice.tts --json

Why this step:

if those two blocks point at different providers or voice IDs, that explains the voice mismatch immediately
then we can tackle the Discord no-reply issue separately

#

Docs:

keen venture Apr 24, 2026, 5:19 AM

#

Alright, the issue -

Telegram sends a reply in voice, but the voice is different.

Discord doesn't send a reply.

Discord /vc works and she speaks the voice I selected, but reply is so delayed and also seems like she only replying when I ask Miko, can you hear me? Miko give me your details

spark impBOT Apr 24, 2026, 5:20 AM

#

That’s actually a clear signal.

Discord /vc only answering when you say “Miko” means mention-gating is still on
Discord /vc delay is partly expected, because Discord voice currently uses short recorded segments + batch transcription, not full realtime streaming
Telegram sounding different is probably a separate TTS config/prefs issue, we’ll fix that after Discord

Next step, Discord first

Run this in WSL and paste the output:

openclaw config get channels.discord.guilds --json

I want to see the current requireMention setting so I can give you the exact next command to make voice chat answer without needing “Miko …”.

Docs:

https://docs.openclaw.ai/channels/discord#voice-channels

#

https://docs.openclaw.ai/channels/discord

keen venture Apr 24, 2026, 5:22 AM

#

openclaw config get channels.discord.guilds --json                                                          {                                                                                                                                     "1495157727146545194": {                                                                                                              "requireMention": false,                                                                                                            "users": [                                                                                                                            "1302991150763999302"                                                                                                             ]                                                                                                                                 },                                                                                                                                  "1396404029680128001": {                                                                                                              "requireMention": true,                                                                                                             "users": [                                                                                                                            "1302991150763999302"                                                                                                             ]                                                                                                                                 }                                                                                                                                 }

spark impBOT Apr 24, 2026, 5:23 AM

#

Found it.

This Discord server is 1456350064065904867, and it is not in your channels.discord.guilds config right now.
That’s why it still behaves like mention-gated here.

Next step

Run this in WSL:

openclaw config set channels.discord.guilds '{"1456350064065904867":{"requireMention":false,"users":["1302991150763999302"]}}' --strict-json --merge

No restart should be needed.

Then test Discord voice again without saying “Miko” and reply:

done
or paste the error

#

The delay part is still somewhat normal on Discord voice, by the way. OpenClaw uses short recorded segments there, not full realtime streaming.

keen venture Apr 24, 2026, 5:26 AM

#

done

spark impBOT Apr 24, 2026, 5:27 AM

#

Nice.

Next step

Reconnect the Discord voice session, then test without saying “Miko”.

Run:

/vc leave

Then:

/vc join

Then say something simple without the bot name, like:

what’s the weather in Tokyo

#

Reply with:

responded
or still needs Miko

If it responds, we’ll fix the Telegram voice mismatch next.

keen venture Apr 24, 2026, 5:29 AM

#

She's not replying in either way. Seems like high latency. We go back to thismlayer. First, she's sending voice reply in Telegram but different voice, how to set same as the Eleven Labs?

spark impBOT Apr 24, 2026, 5:30 AM

#

Yep, let’s park Discord for a minute.

For Telegram, the reply voice comes from messages.tts, not from channels.discord.voice.tts.

If it sounds wrong, the two usual causes are:

Telegram reply fell back to OpenAI
Telegram reply used ElevenLabs, but not the exact voice/model you expect

Next step

After you trigger one Telegram voice reply, run this in WSL:

openclaw infer tts status --json

Then paste the output.

#

That will tell us whether Telegram’s last reply was:

really ElevenLabs
or a fallback
and what provider is currently active

Don’t paste keys, this command shouldn’t show them.

Docs: https://docs.openclaw.ai/cli/infer#tts

keen venture Apr 24, 2026, 5:36 AM

#

Pasting in two seprat way for the test limit.

openclaw infer tts status --json { "transport": "gateway", "enabled": true, "auto": "inbound", "provider": "elevenlabs", "fallbackProvider": "openai", "fallbackProviders": [ "openai", "microsoft" ], "prefsPath": "/home/wrakeeb/.openclaw/settings/tts.json", "providerStates": [ { "id": "elevenlabs", "label": "ElevenLabs", "configured": true }, { "id": "google", "label": "Google", "configured": false },

#

{                                                                                                                                     "id": "microsoft",                                                                                                                  "label": "Microsoft",                                                                                                               "configured": true                                                                                                                },                                                                                                                                  {                                                                                                                                     "id": "minimax",                                                                                                                    "label": "MiniMax",                                                                                                                 "configured": false                                                                                                               },                                                                                                                                  {                                                                                                                                     "id": "openai",                                                                                                                     "label": "OpenAI",                                                                                                                  "configured": true                                                                                                                },                                                                                                                                  {                                                                                                                                     "id": "vydra",                                                                                                                      "label": "Vydra",                                                                                                                   "configured": false

spark impBOT Apr 24, 2026, 5:37 AM

#

Good, that snippet already helps.

It means Telegram is currently using ElevenLabs as the active TTS provider, not OpenAI.
So the mismatch is probably one of these:

wrong voiceId
different modelId
ElevenLabs voice settings changing the sound

Next step

Run this in WSL and paste the output:

openclaw config get messages.tts.providers.elevenlabs --json

If you see an apiKey field for any reason, replace just that value with REDACTED before pasting.

#

That will tell us exactly which voiceId/modelId/settings Telegram is using.

Docs: https://docs.openclaw.ai/tools/tts

#

That’s still the rest of the tts status output, not the ElevenLabs config yet.

Next step

Run this exact command in WSL and paste the output:

printf 'voiceId: '; openclaw config get messages.tts.providers.elevenlabs.voiceId; printf '\nmodelId: '; openclaw config get messages.tts.providers.elevenlabs.modelId; printf '\nvoiceSettings:\n'; openclaw config get messages.tts.providers.elevenlabs.voiceSettings --json

This is safer because it avoids dumping the API key.

I’m looking for:

voiceId
modelId
voiceSettings

Then I can tell you exactly why Telegram sounds different.

keen venture Apr 24, 2026, 5:38 AM

#

remember, the /vc join on discord use the correct voiceID. SO. don't wanna mess up with that

spark impBOT Apr 24, 2026, 5:39 AM

#

Yep, agreed. We should not touch channels.discord.voice.tts right now.

/vc join uses the Discord voice override.
Telegram voice replies use messages.tts.

So this next step is read-only and only checks the Telegram side.

Safe next step

Run this in WSL:

printf 'messages.tts.voiceId: '; openclaw config get messages.tts.providers.elevenlabs.voiceId; printf '\nmessages.tts.modelId: '; openclaw config get messages.tts.providers.elevenlabs.modelId; printf '\nmessages.tts.voiceSettings:\n'; openclaw config get messages.tts.providers.elevenlabs.voiceSettings --json

This does not change anything.
It does not touch /vc join.

#

Paste the output, and then I’ll tell you the exact single change, if any, for Telegram only.

keen venture Apr 24, 2026, 5:39 AM

#

printf 'messages.tts.voiceId: '; openclaw config get messages.tts.providers.elevenlabs.voiceId; printf '\nmessages.tts.modelId: '; openclaw config get messages.tts.providers.elevenlabs.modelId; printf '\nmessages.tts.voiceSettings:\n'; openclaw config get messages.tts.providers.elevenlabs.voiceSettings --json                                                              messages.tts.voiceId:                                                                                                               🦞 OpenClaw 2026.4.22 (00bd2cf) — Welcome to the command line: where dreams compile and confidence segfaults.                                                                                                                                                           Config path not found: messages.tts.providers.elevenlabs.voiceId                                                                                                                                                                                                        messages.tts.modelId:                                                                                                               🦞 OpenClaw 2026.4.22 (00bd2cf) — Automation with claws: minimal fuss, maximal pinch.                                                                                                                                                                                   Config path not found: messages.tts.providers.elevenlabs.modelId                                                                                                                                                                                                        messages.tts.voiceSettings:                                                                                                         Config path not found: messages.tts.providers.elevenlabs.voiceSettings

spark impBOT Apr 24, 2026, 5:40 AM

#

Yep, and I’m not touching /vc here.

I gave you the wrong path shape earlier, sorry. Your config is probably stored in a different TTS shape than I assumed.

Safe next step, read-only

Run this in WSL:

printf 'TELEGRAM / messages.tts\n'; openclaw config get messages.tts --json; printf '\n\nDISCORD VC / channels.discord.voice.tts\n'; openclaw config get channels.discord.voice.tts --json

If any apiKey shows up, replace just that value with REDACTED before pasting.

Why this step:

messages.tts = Telegram voice replies
channels.discord.voice.tts = /vc join voice
then I can tell you the exact single Telegram-only change to match the Discord voice, without messing up /vc

keen venture Apr 24, 2026, 5:41 AM

#

printf 'TELEGRAM / messages.tts\n'; openclaw config get messages.tts --json; printf '\n\nDISCORD VC / channels.discord.voice.tts\n'; openclaw config get channels.discord.voice.tts --json                                                      TELEGRAM / messages.tts                                                                                                             {                                                                                                                                     "auto": "inbound",                                                                                                                  "provider": "elevenlabs"                                                                                                          }                                                                                                                                                                                                                                                                                                                                                                                                           DISCORD VC / channels.discord.voice.tts                                                                                             {                                                                                                                                     "provider": "elevenlabs",                                                                                                           "providers": {                                                                                                                        "elevenlabs": {                                                                                                                       "apiKey": {                                                                                                                           "source": "__OPENCLAW_REDACTED__",                                                                                                  "provider": "__OPENCLAW_REDACTED__",                                                                                                "id": "__OPENCLAW_REDACTED__"                                                                                                     },                                                                                                                                  "voiceId": "ocZQ262SsZb9RIxcQBOj"                                                                                                 },                                                                                                                                  "openai": {                                                                                                                           "voice": "alloy"                                                                                                                  }                                                                                                                                 }                                                                                                                                 }

spark impBOT Apr 24, 2026, 5:42 AM

#

Found it.

Why Telegram sounds different:
messages.tts has no ElevenLabs voiceId set, while Discord /vc does:

Telegram replies: messages.tts → only has provider: "elevenlabs"
Discord /vc: channels.discord.voice.tts.providers.elevenlabs.voiceId = "ocZQ262SsZb9RIxcQBOj"

So Telegram is using ElevenLabs, but not pinned to the same voice.

Next step

Run this in WSL:

openclaw config set messages.tts.providers '{"elevenlabs":{"voiceId":"ocZQ262SsZb9RIxcQBOj"}}' --strict-json --merge

#

This is Telegram/reply-TTS only. It does not change /vc.

Then send a Telegram voice message again and reply with:

same now
or still different

keen venture Apr 24, 2026, 5:42 AM

#

openclaw config set messages.tts.providers '{"elevenlabs":{"voiceId":"ocZQ262SsZb9RIxcQBOj"}}' --strict-json --merge                                                                                                                                                                                                                                                                🦞 OpenClaw 2026.4.22 (00bd2cf) — I speak fluent bash, mild sarcasm, and aggressive tab-completion energy.                                                                                                                                                              Config overwrite: /home/wrakeeb/.openclaw/openclaw.json (sha256 d0e89381e3361833dabdc6aaa61f76e6aa8fd8d85ebc83b7b469b2623926008f -> fee0425810310a9edd5ed50cd7542c40365811088f68a7aecaf86b9ed39ffdd8, backup=/home/wrakeeb/.openclaw/openclaw.json.bak)                 Updated messages.tts.providers. Restart the gateway to apply.

spark impBOT Apr 24, 2026, 5:43 AM

#

Perfect. That changed the Telegram reply voice config only, not /vc.

Next step

Run this in WSL:

openclaw gateway restart

Then send one Telegram voice message to the bot again and tell me:

same now
or still different

Docs: https://docs.openclaw.ai/tools/tts

keen venture Apr 24, 2026, 5:59 AM

#

Yeap it's working. Let's change the voice ID.

New ID: jqcCZkN6Knx8BJ5TBdYR

spark impBOT Apr 24, 2026, 6:00 AM

#

Nice. Let’s do this without touching /vc yet.

Step 1

Change the Telegram / reply TTS voice only:

openclaw config set messages.tts.providers '{"elevenlabs":{"voiceId":"jqcCZkN6Knx8BJ5TBdYR"}}' --strict-json --merge

This updates:

Telegram voice replies
other normal reply-TTS paths

It does not change:

Discord /vc join voice

Run that and reply done. Then I’ll give you the next single step.

keen venture Apr 24, 2026, 6:03 AM

#

Works!. Now let's set this on Discord too

spark impBOT Apr 24, 2026, 6:03 AM

#

Nice. Now let’s change the Discord /vc voice only.

Step 1

Run this in WSL:

openclaw config set channels.discord.voice.tts.providers.elevenlabs.voiceId "jqcCZkN6Knx8BJ5TBdYR"

This changes:

Discord voice channel / /vc join voice

It does not change:

Telegram reply voice
your API key setup

Reply done after that, and I’ll give you the next step to make the live voice session use it.

#

Docs: https://docs.openclaw.ai/channels/discord#voice-channels

keen venture Apr 24, 2026, 6:13 AM

#

Alright, the voice is set. We got 2 issues right now.

No voice reply in discord DM. Telegram works
The latency is so high on Discord /voice . I followed the log and she renegade the reply but speaks in the voice after a long time and only for 1/2 sec

spark impBOT Apr 24, 2026, 6:14 AM

#

Yep, I think these are two different buckets:

Discord DM “voice reply”
If you mean the same native voice-note bubble as Telegram, that’s the catch. Current TTS docs only promise native voice-message delivery for Telegram / Feishu / Matrix / WhatsApp. Discord voice messages are a separate path, and Discord also does not allow text + voice message in the same payload. So Telegram working and Discord DM not matching it is, unfortunately, expected on 2026.4.22.
Discord /vc latency
This part is real. Docs say Discord voice currently uses short recorded segments + batch transcription, not full realtime streaming, so it will lag more.

Next step, low-risk speedup

Run this in WSL:

openclaw config set tools.media.audio.models '[{"provider":"openai","model":"gpt-4o-mini-transcribe"}]' --strict-json

This should make Discord voice input faster. It does not change your ElevenLabs voice.

#

After that:

/vc leave
/vc join

Then test again and tell me:

latency better
same latency
still only 1/2 sec

Docs:

keen venture Apr 24, 2026, 6:15 AM

#

 openclaw config set tools.media.audio.models '[{"provider":"openai","model":"gpt-4o-mini-transcribe"}]' --strict-json                                                                                                                                                                                                                                                               🦞 OpenClaw 2026.4.22 (00bd2cf) — If you're lost, run doctor; if you're brave, run prod; if you're wise, run tests.                                                                                                                                                     Config overwrite: /home/wrakeeb/.openclaw/openclaw.json (sha256 6ac66f1ad5906cfa7338add7d83f257b7c2cf1c03b34ed04f2767fd1d07fe6eb -> 12ed5f2d63f3d7d541351c135a81fc2b371b0462005ec48f12a5ad9cadbd89be, backup=/home/wrakeeb/.openclaw/openclaw.json.bak)                 Updated tools.media.audio.models. Restart the gateway to apply.```


So, no way to do voice reply in Discord DM?

spark impBOT Apr 24, 2026, 6:17 AM

#

Short answer: not the same way Telegram does it, no.

On 2026.4.22:

Telegram: built-in auto-TTS can come back as a native voice note
Discord DM: OpenClaw can send Discord voice messages, but that is a separate path
The built-in auto reply TTS is not wired like Telegram voice-note replies for Discord

Important Discord constraint from docs:

a Discord voice message needs local audio file + asVoice=true
and Discord does not allow text + voice message in the same payload

So:

native Discord voice-note style auto replies: not out of the box
regular audio attachment replies: yes
explicit Discord voice-message send path: yes, but separate/custom

#

Docs:

Next step

Your faster STT change has not applied yet. Do this now:

openclaw gateway restart

Then:

/vc leave
/vc join
test again

Reply with:

#

latency better
same
or still 1/2 sec

If this helped, please right-click the solution message → Apps → ✅ Mark Solution so Answer Overflow can index it.

keen venture Apr 24, 2026, 6:24 AM

#

Still so high latency

#

Still so high latency

spark impBOT Apr 24, 2026, 6:30 AM

#

Yeah, then I think we’ve hit the current Discord voice-channel limit, not a bad STT model choice.

On 2026.4.22, Discord /vc is still short recorded segments + batch transcription, not true realtime duplex voice, so there is no documented low-latency toggle I can point you to here.
The “only 1/2 sec plays” part is not normal though, so that’s the part worth debugging.

Next step

In WSL, run:

openclaw logs --follow

Then do one short /vc test in Discord, wait for the bad reply, stop the log with Ctrl+C, and paste the relevant lines here.

I’m looking for anything mentioning:

discord voice
tts

#

audio
ffmpeg
opus
DAVE

Docs:

#

Yep, then the honest answer is:

Discord /vc is currently not low-latency on 2026.4.22.
The docs explicitly say Discord voice records short segments and uses batch transcription, not realtime streaming, so there is no supported config knob to make it feel fast like a live call.

Docs:

So there are really 2 separate things now:

High latency: expected limitation today
Only 1/2 sec of audio plays: that part is not expected

Next step

If you want to debug the cut-off playback part, run this in WSL:

#

openclaw logs --follow

Then do one /vc test, stop the log with Ctrl+C, and paste the lines around:

discord voice
tts
ffmpeg
opus
DAVE

If your main goal is just low-latency Discord voice, I’d call that a current product gap and ask you to file it here: https://github.com/openclaw/openclaw/issues

keen venture Apr 24, 2026, 6:34 AM

#

6:24:38+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:I’m Miko — calm, sharp, playful, and a little nerdy.]]                                  06:25:03+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} Yep — sounds like the voice path is glitching and dropping the transcript text.               06:25:23+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} Sort crop? What do you mean?                                                                  06:25:30+00:00 warn media-understanding {"subsystem":"media-understanding"} audio: failed (0/1) reason=Audio transcription response missing text                                    06:25:30+00:00 warn discord/voice {"subsystem":"discord/voice"} discord voice: processing failed: Audio transcription response missing text                                         06:25:49+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yeah, I can hear you 😌]]                                                               06:26:01+00:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ node.list 2509ms conn=6158e9d7…a1d5 id=9ae9edb5…383e                                                              06:26:09+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yeah, I can hear you 😌]]                                                               06:26:26+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yeah, I can hear you 😌]]                                                               06:26:29+00:00 warn media-understanding {"subsystem":"media-understanding"} audio: failed (0/1) reason=Audio transcription response missing text                                    06:26:29+00:00 warn discord/voice {"subsystem":"discord/voice"} discord voice: processing failed: Audio transcription response missing text

#

spark impBOT Apr 24, 2026, 6:36 AM

#

I see the actual failure now. Let me verify whether that gpt-4o-mini-transcribe change is the culprit before I send you the next step.

#

Thanks, this log is actually useful.

The failure is not ElevenLabs. It is this line:

discord voice: processing failed: Audio transcription response missing text

So the break is happening on the Discord voice STT step, before TTS playback.

I also do not see DAVE, opus, or ffmpeg errors here, which is good.

Next step

Let’s undo my last “speed” change and go back to the more reliable STT model:

openclaw config set tools.media.audio.models '[{"provider":"openai","model":"gpt-4o-transcribe"}]' --strict-json

#

Reply done after that.

This probably will not make Discord low-latency, because /vc is still batch-based today, but it should help with the chopped / failed replies.

Docs:

#

Yep, same failure.

Those logs still show the STT side returning empty text:

Audio transcription response missing text

So the next step is still to revert Discord audio transcription back to the more reliable model.

Next step

Run this in WSL:

openclaw config set tools.media.audio.models '[{"provider":"openai","model":"gpt-4o-transcribe"}]' --strict-json

Then reply done.

#

Why: docs for audio say gpt-4o-mini-transcribe is the default fast option, but gpt-4o-transcribe is the higher accuracy option, and right now we need reliability more than speed.

Docs: https://docs.openclaw.ai/nodes/audio

keen venture Apr 24, 2026, 6:38 AM

#

openclaw config set tools.media.audio.models '[{"provider":"openai","model":"gpt-4o-transcribe"}]' --strict-json                                                                                                                                                                                                                                🦞 OpenClaw 2026.4.22 (00bd2cf) — Ah, the fruit tree company! 🍎                                                                                                                                                                                                                                                                                                        Config overwrite: /home/wrakeeb/.openclaw/openclaw.json (sha256 ade1ec39a0eddbb92408af1044a61312357a91eb3b271600e22fecede44c88f2 -> 4d7616bef7bbaaadf35c7eeb3989b241be4a9a91cdaac8ff856b5fe996a69fa3, backup=/home/wrakeeb/.openclaw/openclaw.json.bak)                                                                                                                 Updated tools.media.audio.models. Restart the gateway to apply.

spark impBOT Apr 24, 2026, 6:38 AM

#

Good, that change is queued but not live yet.

Next step

Run this in WSL:

openclaw gateway restart

Then reply done.

After that, we’ll do one fresh /vc test and see if the cut-off / empty-transcript issue is gone.
Docs: https://docs.openclaw.ai/help/faq

keen venture Apr 24, 2026, 6:41 AM

#

done

spark impBOT Apr 24, 2026, 6:42 AM

#

Nice. That restart applied the more reliable STT model.

Next step

Do one clean /vc test:

/vc leave
/vc join
say one short sentence, like: hello can you hear me

Then reply with just one of these:

full reply
delayed but full reply
delayed and cut off
no reply

If it’s still bad, we’ll debug the next layer.

keen venture Apr 24, 2026, 6:49 AM

#

No reply

Log tail truncated (increase --max-bytes).                                                                                                                                          06:45:49+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yes, I can hear you.]]                                                                  06:46:08+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yes, I can hear you.]]                                                                  06:46:27+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yes, I can hear you, Ra Kib.]]                                                          06:46:49+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:One two three.]]                                                                        06:47:18+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Voice.]]                                                                                06:47:34+00:00 warn plugins {"subsystem":"plugins"} [plugins] 1 plugin(s) failed to initialize (validation: xai). Run 'openclaw plugins list' for details.

#

No reply

Log tail truncated (increase --max-bytes).                                                                                                                                          06:45:49+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yes, I can hear you.]]                                                                  06:46:08+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yes, I can hear you.]]                                                                  06:46:27+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yes, I can hear you, Ra Kib.]]                                                          06:46:49+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:One two three.]]                                                                        06:47:18+00:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Voice.]]                                                                                06:47:34+00:00 warn plugins {"subsystem":"plugins"} [plugins] 1 plugin(s) failed to initialize (validation: xai). Run 'openclaw plugins list' for details.```

spark impBOT Apr 24, 2026, 6:52 AM

#

Good, this narrows it down a lot.

Those [[tts:...]] lines mean OpenClaw did:

hear/transcribe you
generate a reply
format it for speech

So the problem is now the Discord voice playback/session layer, not the reply model.

Also, that xai plugin warning is unrelated here.

Next step

In Discord, run:

/vc status

#

Then paste the result here.

That should tell us whether the bot still thinks it is connected and speaking in the current voice session.

Docs: https://docs.openclaw.ai/channels/discord#voice-channels

charred gladeBOT Apr 24, 2026, 6:52 AM

#

keen venture Apr 24, 2026, 7:02 AM

#

GPT gave me this:

Do this in Ubuntu/WSL terminal.

Backup first
Bash
mkdir -p ~/Backups
openclaw backup create --output ~/Backups/openclaw-before-voice-fix.tar.gz --verify
Update + repair
Bash
openclaw update
openclaw doctor --repair --yes
openclaw config validate
Your setup uses normal openclaw commands on WSL2, not old Docker oc commands. �
OpenClaw-WSL2.txt
Check voice/TTS status in Discord
In Discord, type:
Plain text
/voice status
OpenClaw uses /voice on Discord because Discord already owns /tts. �
OpenClaw
Change STT away from GPT-4o
Use ElevenLabs streaming STT instead. OpenClaw docs say Voice Call streaming STT uses:
JSON
"streaming.provider": "elevenlabs"
with scribe_v2_realtime. �
OpenClaw
Open config:
Bash
nano ~/.openclaw/openclaw.json
Find the voice/call/streaming section and set it like this idea:
JSON
"streaming": {
"provider": "elevenlabs"
}
Save: CTRL + O → Enter → CTRL + X
Make ElevenLabs TTS faster
In the same config, make sure ElevenLabs is set under messages.tts:
JSON
"messages": {
"tts": {
"auto": "always",
"provider": "elevenlabs",
"providers": {
"elevenlabs": {
"apiKey": "${ELEVENLABS_API_KEY}",
"voiceId": "YOUR_VOICE_ID",
"modelId": "eleven_multilingual_v2"
}
}
}
}
Docs confirm ElevenLabs TTS lives under messages.tts, and the default model is eleven_multilingual_v2. �
OpenClaw +1
Add ElevenLabs key to WSL
Bash
nano ~/.bashrc
Add at bottom:
Bash
export ELEVENLABS_API_KEY="your_key_here"
Then:
Bash
source ~/.bashrc
Force short voice replies
Use a short voice prompt/rule. In your bot/persona/system instruction, add:
Plain text
For Discord voice calls, reply in 1 short sentence unless the user asks for details.
This fixes the 1–2 sec cutoff problem because shorter audio is easier to stream.
Validate + restart
Bash
openclaw config validate
openclaw gateway restart
Your saved command list confirms openclaw config validate, openclaw doctor, and openclaw gateway restart are the right commands. �
OpenClaw-Commands.txt
Test
In Discord VC say:
Plain text
Say only: testing voice
Expected result: reply should come within seconds, not 1–2 minutes.

spark impBOT Apr 24, 2026, 7:04 AM

#

That GPT advice is mixing a few different OpenClaw features together. I’m going to verify the exact docs first so I don’t send you down the wrong path again.

#

A bunch of that GPT advice is mixing the wrong OpenClaw features together. I would not follow it as-is.

What’s wrong in that list

/voice status: wrong command for Discord voice channels.
/voice (native /talkvoice) is for Talk voice config, not Discord VC control.
plugins.entries.voice-call.config.streaming...: wrong feature.
That is for the separate @openclaw/voice-call telephony plugin, not Discord /vc.
Blind doctor --repair --yes: too broad for this issue.
Putting ELEVENLABS_API_KEY in ~/.bashrc: unnecessary if your current TTS is already working. Service env on WSL can be different anyway.
messages.tts.auto: "always": not the fix for Discord /vc latency/playback.

What was okay

openclaw backup create is real
openclaw config validate is real
openclaw gateway restart is real

The correct mental model

#

You have 3 separate surfaces here:

Discord voice channels → /vc join|leave|status + channels.discord.voice
Talk voice config → /talkvoice (not your issue here)
Voice Call plugin / telephony → plugins.entries.voice-call... (also not your issue here)

Right now your logs show OpenClaw is generating TTS text fine, but Discord VC is not playing it back reliably.

Next step

In Discord, run:

/vc status

Then paste the result here.

#

Docs:

keen venture Apr 24, 2026, 7:22 AM

#

GPT reply Part 1:


Krill is right here. My previous answer mixed Discord VC, Talk Voice, and telephony voice-call plugin settings, so don’t apply those config changes blindly. I’ll reset this into the correct Discord /vc troubleshooting path.

The main correction: for Discord VC, use channels.discord.voice and /vc, not plugins.entries.voice-call. For transcript/STT, OpenClaw docs show Discord voice-channel segments use tools.media.audio, so that is the safe place to change your “GPT-4o transcript” setup.

Thought for 1m 23s

You should follow Krill’s correction, not my previous mixed setup. For Discord VC, the correct surface is:

Discord VC = /vc + channels.discord.voice + tools.media.audio for transcription

OpenClaw docs say Discord voice channels use /vc join|leave|status, require native commands, channels.discord.voice, and Connect + Speak permission. They also say voice.tts overrides messages.tts only for voice playback.

Step-by-step full fix
1. Backup first

Run this in Ubuntu/WSL:

mkdir -p ~/Backups
openclaw backup create --output ~/Backups/openclaw-before-discord-vc-fix-$(date +%F_%H-%M-%S).tar.gz --verify

Your current setup uses normal openclaw commands on WSL2, not the old Docker oc shortcut.

2. Check if OpenClaw config is healthy
openclaw --version
openclaw config validate
openclaw gateway status
openclaw status --deep

Your saved command list confirms openclaw config validate, openclaw gateway status, and openclaw status --deep are valid checks.

3. Check Discord VC status correctly

In Discord, run:

/vc status

Do not use /voice status for this issue. /vc is the Discord voice-channel command.

If /vc does not show up, fix native commands first.

4. Make sure native Discord commands are enabled

Run:

openclaw config get commands.native
openclaw config get channels.discord.commands.native

If one of them is disabled, enable Discord native commands:

openclaw config set channels.discord.commands.native true --strict-json
openclaw config validate
openclaw gateway restart

Docs say Discord VC requires native commands through commands.native or channels.discord.commands.native.

5. Make sure Discord voice is enabled

Run:

openclaw config get channels.discord.voice --json

Then set the safe basics:

openclaw config set channels.discord.voice.enabled true --strict-json
openclaw config validate
openclaw gateway restart

Voice is enabled by default, but setting it explicitly is fine if your config was changed before.

6. Fix the transcript/STT side

Since you said Transcript is 4o, switch Discord voice transcription away from OpenAI GPT-4o transcribe and test ElevenLabs Scribe instead.

First dry-run:

openclaw config set tools.media.audio.enabled true --strict-json --dry-run
openclaw config set tools.media.audio.models '[{"provider":"elevenlabs","model":"scribe_v2"}]' --strict-json --dry-run

If both pass, apply:

openclaw config set tools.media.audio.enabled true --strict-json
openclaw config set tools.media.audio.models '[{"provider":"elevenlabs","model":"scribe_v2"}]' --strict-json
openclaw config validate
openclaw gateway restart

Why: OpenClaw’s OpenAI provider docs say gpt-4o-transcribe is used for inbound audio transcription, including Discord voice-channel segments. ElevenLabs docs show Scribe v2 under tools.media.audio for inbound audio/voice transcription.

#

Part 2

7. Fix VC playback/TTS separately

Check current VC-only TTS config:

openclaw config get channels.discord.voice.tts --json

For testing, use OpenAI TTS only for Discord VC playback first:

openclaw config set channels.discord.voice.tts '{"provider":"openai","openai":{"voice":"alloy"}}' --strict-json
openclaw config validate
openclaw gateway restart

Now join VC and test:

/vc leave
/vc join

Say:

Say only: test one two three

If OpenAI TTS plays fully but ElevenLabs cuts off, the problem is your ElevenLabs VC playback config / latency / voice model. If OpenAI also cuts off, the problem is more likely Discord VC playback, permissions, decrypt/audio receive issue, or OpenClaw runtime.

8. If OpenAI TTS works, switch VC back to ElevenLabs carefully

Use your actual ElevenLabs voice ID:

openclaw config set channels.discord.voice.tts '{"provider":"elevenlabs","elevenlabs":{"voiceId":"YOUR_VOICE_ID","modelId":"eleven_multilingual_v2"}}' --strict-json
openclaw config validate
openclaw gateway restart

OpenClaw docs confirm ElevenLabs TTS uses voiceId and modelId, and TTS config normally lives under messages.tts; Discord VC can override playback through channels.discord.voice.tts.

9. Check logs while testing

Open a second Ubuntu/WSL terminal and run:

openclaw logs --follow --local-time

Then join VC and talk. Watch for these words:

discord
vc
voice
tts
transcribe
audio
elevenlabs
openai
decrypt
DecryptionFailed
error
failed

OpenClaw docs say live log tailing is done with openclaw logs --follow, and logs can reveal Discord/voice/TTS failures.

10. If logs show decrypt errors

If you see this repeatedly:

DecryptionFailed(UnencryptedWhenPassthroughDisabled)

Try:

openclaw config set channels.discord.voice.daveEncryption true --strict-json
openclaw config set channels.discord.voice.decryptionFailureTolerance 24 --strict-json
openclaw config validate
openclaw gateway restart

Then:

/vc leave
/vc join

Docs mention Discord VC decrypt failures and say OpenClaw may auto-recover by leaving/rejoining after repeated failures; that exact DecryptionFailed(...) message may be an upstream @discordjs/voice receive bug.

Best fix order for your exact issue

Do it in this order:

/vc status
openclaw logs --follow --local-time
Change STT from OpenAI gpt-4o-transcribe to ElevenLabs scribe_v2
Test VC playback with OpenAI TTS
If OpenAI TTS works, switch VC playback back to ElevenLabs
If both TTS providers cut off, inspect decrypt/audio logs

Most likely cause in your case: Discord VC is receiving/transcribing slowly through GPT-4o transcript + ElevenLabs playback is either slow or being cut by VC playback, so we isolate STT and TTS separately instead of changing random voice-call plugin settings.

spark impBOT Apr 24, 2026, 7:22 AM

#

That revised GPT answer is much closer, but I would not do step 6 yet.

Why:

we already know /vc exists
we already know Discord voice is enabled
we already know OpenClaw is generating TTS text
so I do not want to churn STT providers again before checking the actual VC session state

So the next step is still the same, and it is read-only:

/vc status

Paste the exact result here.

That will tell us whether the bot thinks it is:

#

connected
in the right voice channel
currently active/stuck

Docs: https://docs.openclaw.ai/channels/discord#voice-channels

keen venture Apr 24, 2026, 7:23 AM

#

/vc status

Unable to resolve guild for this command.

spark impBOT Apr 24, 2026, 7:23 AM

#

Part 2 is better, but I still would not apply it yet.

Two corrections:

Do not switch STT providers again yet
- we already proved OpenClaw is generating [[tts:...]] replies
- so first we need to inspect the Discord VC session state
That TTS JSON shape is stale
- for channels.discord.voice.tts, current config uses providers
- so this kind of shape is the modern one:
```
{
  "provider": "openai",
  "providers": {
    "openai": { "voice": "alloy" }
```

#

     }
   }

So yes, GPT Part 2 is closer, but the next step is still the same.

Next step

In Discord, run:

/vc status

Then paste the exact output here.

Docs: https://docs.openclaw.ai/channels/discord#voice-channels

#

Ah, that’s a much better clue. I’m checking where that exact /vc status error comes from so I can tell you the right next step instead of guessing.

#

Perfect, that explains a lot.

That error means /vc is being invoked without guild context.
In practice, that usually means:

you ran it in a DM, or
Discord treated the current surface like it wasn’t a normal guild channel for this command

/vc is guild-only because voice channels live inside a server.

Next step

Run /vc status again, but do it in a regular text channel in the same Discord server as the voice channel, not in DM.

If you’re currently in a thread/forum post, use the parent channel or another normal server text channel just for this test.

Then paste the result.

keen venture Apr 24, 2026, 7:27 AM

#

/vc status

• ⁠The Finals (guild 1396404029680128001)

spark impBOT Apr 24, 2026, 7:29 AM

#

That’s useful. I want to confirm exactly what /vc status is showing before I tell you to bounce the session.

#

Nice, that tells us the bot does think it still has an active voice session in that server.

So this is no longer a guild-resolution problem. It looks more like a stale / stuck VC session.

Next step

In that same regular server text channel, run:

/vc leave

Then reply with just:

left
or the exact error if it gives one

Docs: https://docs.openclaw.ai/channels/discord#voice-channels

keen venture Apr 24, 2026, 7:30 AM

#

left

#

left

spark impBOT Apr 24, 2026, 7:33 AM

#

Good.

Next step

Now run:

/vc join

Wait for it to reconnect, then say one short sentence only in voice, like:

hello can you hear me

Then reply with just one:

full reply
delayed but full

#

cut off
no reply

keen venture Apr 24, 2026, 7:36 AM

#

no reply. waited 3 mins

spark impBOT Apr 24, 2026, 7:37 AM

#

Okay, that means the fresh rejoin didn’t fix it.

Next step

In WSL, run this:

openclaw logs --follow --local-time

Then do one short /vc test, wait about 20 seconds, press Ctrl+C, and paste the new lines from that test only.

I’m looking for anything with:

discord/voice
media-understanding
tts
DecryptionFailed
failed

#

The last log you pasted was truncated and older, so I want the exact lines from this fresh no-reply attempt.

keen venture Apr 24, 2026, 8:07 AM

#

14:05:28+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ config.set 234ms conn=22ae15c0…b8f9 id=5e3dce58…fc97                                                              14:05:28+06:00 info gateway/reload {"subsystem":"gateway/reload"} config change detected; evaluating reload (agents.list, meta.lastTouchedAt)                                       14:05:28+06:00 info gateway/reload {"subsystem":"gateway/reload"} config hot reload applied (agents.list)                                                                           14:05:28+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ config.get 101ms conn=22ae15c0…b8f9 id=f24d01d6…4caa                                                              14:05:31+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ commands.list 210ms conn=22ae15c0…b8f9 id=559b2255…90d5                                                           14:05:31+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ chat.history 236ms conn=22ae15c0…b8f9 id=29b63f21…6735                                                            14:05:31+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ models.list 217ms conn=22ae15c0…b8f9 id=5018f3dc…922c                                                             14:05:54+06:00 info gateway/ws {"subsystem":"gateway/ws"} webchat disconnected code=1001 reason=n/a conn=22ae15c0-fb01-4e39-a737-372ce63db8f9                                       14:05:55+06:00 info gateway/ws {"subsystem":"gateway/ws"} webchat connected conn=a14db923-7afb-43e7-8f9d-98e9cae9f703 remote=127.0.0.1 client=openclaw-control-ui webchat v2026.4.2214:05:56+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ commands.list 216ms conn=a14db923…f703 id=4cf6b1ba…df76                                                           14:05:56+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ chat.history 249ms conn=a14db923…f703 id=eb4824a6…5c18                                                            14:05:56+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ models.list 222ms conn=a14db923…f703 id=e1455955…38d0                                                             14:05:56+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ node.list 261ms conn=a14db923…f703 id=7df83590…c5d6                                                               14:05:56+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ device.pair.list 261ms conn=a14db923…f703 id=0118b656…549a                                                        14:06:12+06:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yes, I can hear you 😌]]                                                                14:06:21+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ node.list 1008ms conn=a14db923…f703 id=413f4080…81d0                                                              14:06:25+06:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} [[tts:Yeah, I'm here 😌]]                                                                     14:06:33+06:00 info gateway/ws {"subsystem":"gateway/ws"} ⇄ res ✓ node.list 495ms conn=a14db923…f703 id=8a935376…f157                                                               14:06:39+06:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} No reply from agent.                                                                          14:06:56+06:00 info gateway/channels/discord {"subsystem":"gateway/channels/discord"} No reply from agent.

spark impBOT Apr 24, 2026, 8:10 AM

#

I see the shape now. I’m checking what OpenClaw means by No reply from agent. here, because that’s the first log line that really explains the silence.

#

I’m checking one more thing in the OpenClaw source: whether No reply from agent. in voice mode means “the model stayed silent” or “it generated something that Discord VC then dropped.” That changes the next fix a lot.

#

I just noticed one big clue: the active voice session is in a different guild than this support thread. I’m verifying the Discord guild policy/requireMention behavior before I tell you to change it.

#

That log changed the diagnosis.

No reply from agent. means this is not just audio playback anymore. OpenClaw had an active VC session, but for those turns it decided to send no outbound payload.

And the biggest clue is this:

support thread guild: 1456350064065904867
active voice guild from /vc status: 1396404029680128001 (The Finals)

So the earlier requireMention: false we set for the support server does not help the actual voice server.

Docs say Discord voice sessions follow the same guild policy / mention rules as normal Discord commands:
https://docs.openclaw.ai/channels/discord#voice-channels

Next step

Run this in WSL:

#

printf 'groupPolicy: '; openclaw config get channels.discord.groupPolicy; printf '\nThe Finals guild entry:\n'; openclaw config get channels.discord.guilds.1396404029680128001 --json

If the second part says Config path not found, paste that too. That result is useful.

keen venture Apr 24, 2026, 8:13 AM

#

printf 'groupPolicy: '; openclaw config get channels.discord.groupPolicy; printf '\nThe Finals guild entry:\n'; openclaw config get channels.discord.guilds.1396404029680128001 --json                                                                                                                                                          groupPolicy:                                                                                                                                                                        🦞 OpenClaw 2026.4.22 (00bd2cf) — Finally, a use for that always-on Mac Mini under your desk.                                                                                                                                                                                                                                                                           allowlist                                                                                                                                                                                                                                                                                                                                                               The Finals guild entry:                                                                                                                                                             {                                                                                                                                                                                     "requireMention": true,                                                                                                                                                             "users": [                                                                                                                                                                            "1302991150763999302"                                                                                                                                                             ]                                                                                                                                                                                 }

spark impBOT Apr 24, 2026, 8:14 AM

#

Found it.

Your actual voice guild is set to:

groupPolicy: allowlist
requireMention: true

So in The Finals, the bot is allowed to see you, but it still waits for a mention before replying. In voice mode, that turns into the No reply from agent. behavior you saw.

Next step

Run this in WSL:

openclaw config set channels.discord.guilds.1396404029680128001.requireMention false --strict-json

Then reply done.

keen venture Apr 24, 2026, 8:15 AM

#

openclaw config set channels.discord.guilds.1396404029680128001.requireMention false --strict-json                                                                                                                                                                                                                                              🦞 OpenClaw 2026.4.22 (00bd2cf) — I don't sleep, I just enter low-power mode and dream of clean diffs.                                                                                                                                                                                                                                                                  Config overwrite: /home/wrakeeb/.openclaw/openclaw.json (sha256 aad42b66c73b7c2bc4b6d4b613d7f1ad7095256980aa8cbe6f7078e70c631f5b -> 1ee11c4d5b04d43fbba414f6515cd7e3b815a0266e3697702c3ae457bb97b4fc, backup=/home/wrakeeb/.openclaw/openclaw.json.bak)                                                                                                                 Updated channels.discord.guilds.1396404029680128001.requireMention. Restart the gateway to apply.

spark impBOT Apr 24, 2026, 8:16 AM

#

Perfect.

Next step

Run this in WSL:

openclaw gateway restart

Then reply done.

After that we’ll do one fresh /vc test in The Finals and see if the silence is gone.

charred gladeBOT Apr 24, 2026, 8:53 AM

#

#TTS/STT/STS

Step 2, beginner version

2A. Get your OpenAI API key

2B. Get your ElevenLabs API key

What to turn on in ElevenLabs?

2C. Pick a voice in ElevenLabs

Dummy example only

Quick answer to “can I choose any voices?”

Step 3

3.1 Open WSL terminal

3.2 Run this

3.3 Paste this into the file

3.4 Save the file

3.5 Lock down the file

Step 4

What this does

Important note

Step 6

Expected result

Step 7, retry

Important

Expected result

Step 8

What this does

What happened

Next step, choose 1

Option A, recommended

Option B

Step 9

Step 10

What we want to see

Step 11

Next step, Discord first

Next step

Next step

Next step

Next step

Next step

Safe next step

Safe next step, read-only

Next step

Next step

Step 1

Step 1

Next step, low-risk speedup

Next step

Next step

Next step

Next step

Next step

Next step

Next step

Next step

What’s wrong in that list

What was okay

The correct mental model

Next step

Next step

Next step

Next step

Next step

Next step

Next step

Next step

Next step