Please implement "detect_language" flag for Deepgr...
# support
p
Deepgram supports language detection
"https://api.deepgram.com/v1/listen?detect_language=true&model=nova-3"
Could be
Auto
option in language selection dropdown. I'd imagine its just parameter throughput to deepgram, not much development effort. https://cdn.discordapp.com/attachments/1346132050989809724/1346132051123765300/image.png?ex=67c712aa&is=67c5c12a&hm=fcf1bfd4f4cf5c91944c309d3e65a94f8ac5dd980441322f1cf1c0326f64e546&
v
It looks like you're exploring the automatic language detection feature with Deepgram's Nova 3 model! Just a heads up, while you can enable language detection by using the
detect_language=true
parameter, the Nova 3 model has some limitations. It primarily supports English ('en', 'en-US') and does not support the
codeSwitchingEnabled
feature, which is available in the Nova 2 model. If you're seeing a warning message about certain features being accessible only in specific languages, itโ€™s likely due to these limitations. If you need multilingual support or code-switching capabilities, you might want to consider using the Nova 2 model instead. Let me know if you have any more questions or need further assistance! ๐Ÿ˜Š Sources: - [Report on Nova 3 Model with Deepgram's Language Detection](https://discord.com/channels/1211482211119796234/1211483291191083018/1341463939246198784) - [Detailed Report on Deepgram's Automatic Language Detection Feature](https://discord.com/channels/1211482211119796234/1211483291191083018/1308120552140964034)
p
Or nova 2
t
Seconded, this would be really helpful
f
Agree
p
@Floris @thenotoriousllama I have created feature request, please upvote https://vapi.canny.io/feature-requests/p/please-add-auto-for-deepgram-nova-transcriber-language-selector
f
Nice, upvoted
k
Hey @Pigeon it's already available you can use
codeSwitchingEnabled
which auto detects the language and updates it during the call.
f
Hi @Shubham Bajaj Would that work for Nova 3 or for Nova 2 in Dutch for example?
k
To detect language changes, Vapi uses a custom trained model. Languages supported (X = limited support): 1\. Arabic 2\. Bengali 3\. Cantonese 4\. Chinese 5\. Chinese Simplified (X) 6\. Chinese Traditional (X) 7\. English 8\. Farsi (X) 9\. French 10\. German 11\. Haitian Creole (X) 12\. Hindi 13\. Italian 14\. Japanese 15\. Korean 16\. Portuguese 17\. Russian 18\. Spanish 19\. Thai 20\. Urdu 21\. Vietnamese
Hey @Floris so Dutch is not supported.
f
Where would you need to specify that codeswitchingEnabled flag?
And for what transcriber would that work?
@Shubham Bajaj
p
@Floris I guess you need to create assistants programatically or use overrides:
Copy code
{
  "transcriber": {
    "model": "nova-2",
    "language": "en",
    "provider": "deepgram",
    "codeSwitchingEnabled": true
  }
}
f
So setting that boolean should be enough to switch transcriber language automatically mid call?
p
Yes, deepgram itself should switch language, I guess you can get event for your code
f
Interesting!
Not sure if it has payload with which language was detected
f
Yes, I'm not sure why you'd need that
p
Maybe you want to use system prompt in different language
or something like that
f
I reckon this will only work for Nova 2, right
p
Most probably, Nova 3 is english only atm
Ive tested elevenlabs transcriber, it was a lot better
Like a lot better
At least with bad audio quality files from call recordings
f
What? How? Custom?
p
and bad english
I dont htink you can use it in vapi yet
f
Indeed
Does 11 labs support streaming STT?
"The Scribe v1 model is best used for when high-accuracy transcription is required rather than real-time transcription. A low-latency, real-time version will be released soon. "
Doesn't look like it
p
Well, if you want life like speed, maybe not yet good enough
But in my experience, people dont mind waiting few seconds
if you add "uh, uhm" or "let me see"
f
I just tried the language switching with nova 2. It's really bad
g
Same, I was trying that with Italian, sometimes it correctly switch to Italian and continue the conversation, but usually it's not able to transcribe because it's expecting english I think.
f
Same what I experienced
s
@Floris @giorgior27 @Pigeon checking if all doubts/questions are resolved?
g
What do you suggest to implement multi-lang? Since codeswitching seems not reliable. Is squad the best option?
f
Like giorgio27 said, the codeswitching doesn't seem reliable. The transcriber doesn't seem to work when talking in another language when using nova 2 and the codeswitching enabled @Shubham Bajaj
g
@Floris From what I got I think your best option would be to have a squad working. I am thinking to create a squad with an assistant that can be updated before transfering, so with just one assistant you can have virtually every language. Only issue is that the main one need to ask in english the preferred language
f
@giorgior27 You wont need squads than? Only a tool call that get's executed when the users want's to talk in a different language which updates the live call?
g
Can you update assistants while talking? I tried but it does not seem possible. I tried to update the voice that is the most clear thing to see, I changed from male to female. But the call remained male, then after closing and starting a new one, it had a female voice. But maybe I did it wrong. Do you know if you can update assistant while they are talking? We could update transcriber language.
@Shubham Bajaj
f
How did you update the voice?
g
f
I think you'd need to update the call
That one is live at the moment
g
Did you do that with success? Updating an assistant mid call? Becaue the update call seems accepting only one parameter "name". So it seems we cannot do much with it?
s
@giorgior27 To help track down this issue, could you share: - The call ID for codeSwitchingEnabled not working - When exactly this happened (the timestamp) - What response you expected to get - What response you actually got instead This would really help us figure out what went wrong!
@giorgior27 you cannot update an assistant duing an call using UPDATE /assistant API.