I've tried multiple ways of adding keyterm to my assistant's transcriber via the API but each time I've added some, the call would not work. TO DO TO REPRODUCE : add key term to the assistant via the API For example : "transcriber": { "model": "nova-3", "language": "multi", "numerals": false, "provider": "deepgram", "endpointing": 300, "keyterm": ["Eva", "Jean", "Benoit", "Henry", "assistante", "virtuelle"], "fallbackPlan": { "transcribers": [ { "model": "gemini-2.0-flash", "language": "Multilingual", "provider": "google" } ] } }, **OBSERVED **: call happening without transcription, hence closing after silence timeout. **EXPECTED **: 1. The call to happen normally, deepgram connection working smoothly 2. If not, the transcriber should use the fall back plan and use my other transcriber (Google gemini flash) I have set up. Call id example : e24e6fb3-28e4-497a-8c05-87ba8c513219 From assistant id : 8abc3963-a71e-43e3-8465-ae8568d27179

Hi there! 👋 I see you're having trouble with Deepgram nova-3 keyterm functionality in your Vapi configuration. I'm unable to determine the exact cause of why your fallback mechanism isn't working as expected. However, I can definitely help with the Deepgram nova-3 keyterm issue based on information I've found. Looking at your configuration, I've identified a few potential issues: 1. **Language Setting Issue**: According to the information I found, nova-3 has restricted language support and currently only supports 'en' and 'en-US' - not 'multi'. Your configuration shows:

json
   "language": "multi"

This could be causing the problem as [nova-3 only supports English languages](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1341463939246198784). 2. **Keyterm vs Keywords**: There's an important distinction between these parameters. For Nova-3, you should specifically use

keyterm

(which you are), while Nova-2 uses

keywords

. You've got this part correct, but it's worth confirming. 3. **Potential 502 Error**: Some users have reported that Deepgram on Nova-3 returns 502 errors when certain configurations are used. This might be happening in your case as well. Here are some troubleshooting steps you can try: 1. Change the language setting to 'en' or 'en-US' instead of 'multi' since nova-3 only supports English.

json
   "language": "en"

2. **Check Deepgram Status**: Verify if [Deepgram is experiencing any degradation](https://status.deepgram.com) that might be affecting your integration. 3. **Verify your Vapi configuration**: Double-check your API keys and permissions to ensure they're correct and haven't expired. 4. **Try without keyterms first**: Test your configuration without any keyterms to establish a baseline, then add them back one by one to identify if a specific term is causing issues.

So it works without the keyterm. Error is the following : 17:15:50:059[ERROR] Transcriber failed (reason: Received network error or non-101 status code.) I suspect it's due to the new multilingual language of Deepgram and its keyterm compatibility ?

checking if this is resolved/solved for you?

@Shubham Bajaj no, not solved. Call id in my initial post is still valid. This one can be used as well : 28931920-e1af-49a7-a4c8-b6793a565384. Every time I specify keyterms with nova-3 multi, it fails to launch the transcriber

🔵 17:15:50:059 Deepgram Websocket Error: { "message": "Received network error or non-101 status code." } 🔵 17:15:50:060 Deepgram Transcriber Error: pipeline-error-deepgram-returning-502-network-error Hey Jeebs, This error has been resolved. It was originated from Deepgram side due to a spike in network requests. I'll suggest you try making another call now and let me know how it goes for you.

@Shubham Bajaj this is getting very weird : I can't even set my transcriber to nova-3 anymore (see attached video). Even without keyterm or whatsoever ! Same issue with the UI where I can change to nova-3 but when "getting" the assistant via API it still says nova-2. Assistant id : 7727ea52-ef05-4ad2-ab4b-658b70d3ff36 https://cdn.discordapp.com/attachments/1357042489554632794/1359450670339457044/Nova3-Issue.mp4?ex=67f78699&is=67f63519&hm=3192c06d69129e1bbc84b787f81922e2f258af54033be3a7e1407558104f5e0c&

Got it will get back to you soon

@Jeebs from eva.be looking into it.

@Shubham Bajaj tested both format, but still the error. This call for example I used the right one : ed710336-5123-494b-a62d-5b1765e218b8 Still the same error Or this one (removed the "é" in keyword) : 52ab8c8a-b4d0-4c0c-8d3d-3aaac221c5c7

The first call ended because the customer entered the call.

🔵 14:15:12:447 Initializing Deepgram. Options: {
  "punctuate": true,
  "numbers": true,
  "no_delay": true,
  "encoding": "linear16",
  "channels": 2,
  "multichannel": true,
  "filler_words": true,
  "interim_results": true,
  "language": "multi",
  "model": "nova-3",
  "sample_rate": 44100,
  "mip_opt_out": false,
  "keyterm": ["Florence", "Wéra", "docteur", "Genappe"],
  "endpointing": 300,
  "numerals": false
}...
 "endedReason": "customer-ended-call",

@Shubham Bajaj indeed, I have ended the call because there was no answer 😬 Basically, when there is no transcriber, no sound get captured hence also no answer get voiced by the LLM & voice models. This happened in both calls ; I've ended the call because I knew it wasn't working. In both cases, the logs show an error for the transcriber.

😨 you are right @Shubham Bajaj, their nova-3 is really under-par in the end 😭 Let me know if you plan to integrate the latest gladia model, that seems the best shot for multilingual ! I'll mark this ticket as closed

Sure.