[Bug] WebSocket TTS quality degradation — mispronunciation & abnormal pitch since ~11PM PST Mar 10

[light dove]

Hi Fish Audio team,

We're experiencing a degradation in TTS output quality via the WebSocket API.

Timeline:

Working correctly: ~6:00 PM PST, Mar 10
Issue onset: ~11:00 PM PST, Mar 10

Symptoms:

Incorrect word readings (misread kanji/words)
Abnormal pitch and accent in Japanese TTS output

Isolation testing:

┌────────────────────────────────────────┬──────────────────┐
│ Method │ Result │
├────────────────────────────────────────┼──────────────────┤
│ Fish AI Studio (web) │ Correct │
├────────────────────────────────────────┼──────────────────┤
│ WebSocket API via Pipecat (production) │ Mispronunciation │
├────────────────────────────────────────┼──────────────────┤
│ WebSocket API standalone (no Pipecat) │ Mispronunciation │
└────────────────────────────────────────┴──────────────────┘

Since the issue reproduces when calling the WebSocket endpoint directly without any framework, we believe the root
cause is on the WebSocket API side.

Minimal repro:
git clone https://github.com/yuki901/fish-audio-test.git
cd fish-audio-test
pip install ormsgpack httpx-ws python-dotenv
echo "FISH_AUDIO_API_KEY=<your_api_key>" > .env

python fish_tts_ws.py
"お電話ありがとうございます。お問い合わせ窓口でございます。恐れ入りますが、御社名とお名前をお願いいたします。"
-o output.mp3
--voice 46745543e52548238593a3962be77e3a
--model s2-pro

WebSocket details:

Endpoint: wss://api.fish.audio/v1/tts/live
Protocol: start → text → stop events (ormsgpack-encoded)
Params: format=mp3, sample_rate=44100, latency=balanced, model header s2-pro

Expected: Natural Japanese pronunciation matching Studio output
Actual: Incorrect readings and unnatural pitch/accent

Were any changes made to the WebSocket API around 11 PM PST yesterday? Any ETA on a fix would be appreciated.

Thanks!

Audio comparison (both in repo):
output_studio.mp3 — Generated via Fish AI Studio (correct)
sample_output.mp3 — Generated via WebSocket API (broken)

[violet sky]

they are sharing the same backend, will take a look soon

[light dove]

Thank you! Any progress?

[light dove]

Oh It is fixed. Thank you so much!