river hawk
I'm experiencing really really really high latency with my voice PE, though I check the debug tab in HA the latency looks normal
As you can see from the screenshot, latency is less than a second, but the device seems to pretty consistently take 20-30 seconds to perform an action and then respond
It's been like this consistently for the past week
frigid bluff
Well your google gen AI integration takes 14 seconds to respond. Try reducing number of exposed entities.
river hawk
Could it be an issue with home assistant cloud stt? It looks like it took ~8 seconds for the voice activity detection to finish?
river hawk
Well your google gen AI integration takes 14 seconds to respond. Try reducing nu...
That's 0.14s. and processed locally. Not 14 seconds haha
140 milliseconds
neon root
its the tts as well
you can run a test with the mobile version. my LLM responds practically immediately but they set up piper to do full text conversion rather than sentence by sentence
causes a huge delay for paragraphs etc.
A possible fix would be sentence by sentence and then it would be near insant replies but until that's supported it will be a bottleneck no matter your hardware.
cant speak for cloud tho. im sure there may be some tts / cloud providers that are doing sentence by sentence transcription but any of the local ones - piper kokoro etc are all full responce in then full responce out text to speech
I think piper does support sentence by sentence too but yeah not a thing in HA
frigid bluff
It's not that.
This request has short response, and also it's processed fully locally.
Something is wrong.
neon root
It's not that.
This request has short response, and also it's processed fully lo...
oh yeah a short responce being that late then something is up
river hawk
It's not that.
This request has short response, and also it's processed fully lo...
Same also happens with requests processed via the LLM that aren't handled locally
Ok I think it's the speech to text/VAD
Here are two tests. First is with local faster-whisper
The second with home assistant cloud
And their respective debug info
events:
- type: run-start
data:
pipeline: 01jmytgxhnnekyjdtt7r3ctxqx
language: en
conversation_id: 01JVVBEVC6GPKXRW7H7HJC6EJ6
tts_output:
token: ulJdYK2y2rM6woxKllM5Yw.flac
url: /api/tts_proxy/ulJdYK2y2rM6woxKllM5Yw.flac
mime_type: audio/flac
timestamp: "2025-05-22T06:39:44.939540+00:00"
- type: stt-start
data:
engine: stt.home_assistant_cloud
metadata:
language: en-GB
format: wav
codec: pcm
bit_rate: 16
sample_rate: 16000
channel: 1
timestamp: "2025-05-22T06:39:44.940132+00:00"
- type: stt-vad-start
data:
timestamp: 1190
timestamp: "2025-05-22T06:39:49.709097+00:00"
- type: stt-vad-end
data:
timestamp: 3160
timestamp: "2025-05-22T06:40:21.467609+00:00"
- type: stt-end
data:
stt_output:
text: Turn off the lights.
timestamp: "2025-05-22T06:40:21.495721+00:00"
- type: intent-start
data:
engine: conversation.google_generative_ai_2
language: en-GB
intent_input: Turn off the lights.
conversation_id: 01JVVBEVC6GPKXRW7H7HJC6EJ6
device_id: 065aafe0357846a5b40b1ebeb576e8a7
prefer_local_intents: true
timestamp: "2025-05-22T06:40:21.496018+00:00"
It took 30 seconds for voice activity detection to end
pastel spindle