Has anyone tried scribe v2 realtime stt with their...
# agents
a
Has anyone tried scribe v2 realtime stt with their server-side vad option? It seems really good on paper but im having really bad performance and i dont know if im doing something wrong
to be more specific, i get partial messages like this:
Copy code
DEBUG  livekit.plugins… Received message   {"pid": 704,        
colly-agent  |                                          type               "job_id":           
colly-agent  |                                          partial_transcript "AJ_PyxwyE8sy7cX",  
colly-agent  |                                          : {'message_type': "room_id":          
colly-agent  |                                          'partial_transcrip "RM_ECFxeokEA6d9"}  
colly-agent  |                                          t', 'text':
But it keeps on showing the same partial message and never actually send the text to the llm if that make sense
Found your new https://github.com/livekit/agents/blob/main/examples/other/elevenlab_scribe_v2.py example. Working much better. Thanks for the great docs 🙂
🙌 1
p
Have you compared it with Deepgram? Is it better?
I need to give it a go at some point but haven't had time yet!
c
I just tried, the STT latency is impressive but my transcript stays in partial state for long period of times (like 1s) before being committed which kills the overall latency Maybe I'm not configuring things right, I've tried with server side and livekit VAD.
a
I have the server side vad as well as the silero vad, just as the example suggested and it is working much better, but I can’t perceive a. Big difference between using Deepgram. I’m currently running an ab testing on production to get more data about this
c
I find 11 labs STT is really slow due to the transcript thing. How have you fixed it? @ancient-judge-59849 deepgram is super fast
I tried the example but it takes like 5 seconds for the speech to be committed to the transcript
a
I did not find good performance on it. I think it is optimized to being use on their own agent
c
Yeah it’s pretty bad. Not sure why they marketed it as “real-time” it’s barely that.
Might be the worst STT I’ve seen tbh in terms of both accuracy and speed…which is odd because they make a really good TTS
b
You guys facing these kind if situations? I tried scribe v2 with the param
min_silence_duration_ms
and got forbidden. Works fine without this param though. 🤦
Screenshot 2568-12-04 at 14.30.38.png
a
Havent seen that issue. Did you enable scribe from the elvenlabs interface?
b
The real issue is that
min_slient_duration_ms
should be 50 - 2000, although the ui says default as 2500 🥲. I've also noticed ElevenLab has updated there api response message to be more accurate.