VAPI•2y ago

Transcript doesn't match LLM inputs/outputs

and it's sometimes totally unusable in french.

For example (call id: 4d88522c-65a3-4f0d-84bf-7787a90b9799):
First message of my assistant : "Bonjour, je suis l'IA de Lonestone"
In the transcript : "Bonjour, je suis Lia de Launcetown. Bonjour, je suis Lia de Lone Stone"

Vapi has the exact input, so it should be the same in the transcript.

Ask Inkeep•5/30/24, 5:18 PM

Hi @Godefroy @Lonestone, I'm an AI assistant for Vapi. I'll analyze available sources and get back to you shortly.

Ask Inkeep•5/30/24, 5:18 PM

-----------------------------
While I generate an answer, here are some sources that might be helpful:

Core Models
Web Calling
Orchestration Models
Update Assistant
-----------------------------

Ask Inkeep•5/30/24, 5:18 PM

Answer

It seems like you are experiencing issues with the accuracy of the transcription in French. This discrepancy can arise due to several factors including accents, speech clarity, or background noise which might affect the transcription model's performance support[Final Call Transcript] Why AI is not precise? .

Vapi uses advanced models for transcribing speech, but like all automated systems, it may not always perfectly capture or interpret spoken language, especially when dealing with homophones or similar-sounding phrases in different dialects or languages.

If the transcription errors are significantly impacting the usability, you might consider providing feedback directly on the specific instances where the discrepancies occur. This can help improve the model's accuracy over time. For now, ensuring clear, slow, and distinct pronunciation during calls can sometimes help mitigate these issues.

For further details on how Vapi handles speech-to-text and other operations, you can refer to the official documentation on Orchestration Models (2) .

Sources
[1] [Final Call Transcript] Why AI is not precise?
[2] Orchestration Models

AAsk Inkeep ### Answer It seems like you are experiencing issues with the accuracy of the tr...

Ask InkeepAPP•5/30/24, 5:19 PM

Got it. Tagging @1231325148079194222 for additional help.

living-lavender•5/30/24, 8:24 PM

Use Deepgram's French model. It will fix your issue.

Also, check out this video tutorial: https://www.youtube.com/watch?v=AH5oS2w3EcQ

YouTubeMaximilian

How to create good multilingual voice-AI assistants (Vapi x GPT-4o)

In this video, I show how to create multilingual Voice-AI assistants that are actually good using Vapi and GPT-4o

.

Make.com: https://www.make.com/en/register?pc=maximilian
Vapi.ai: https://vapi.ai/?ref=maximilian

Optimize Streaming Latency (Elevenlabs): https://elevenlabs.io/docs/api-reference/streaming

Email: max@flowficiency.io
Calendly:...

dead-brownOP•5/30/24, 9:22 PM

@Sahil Why would the first message be processed by the STT? (same for any other message from the AI)
Btw Whisper is better than Deepgram in French.

dead-brownOP•6/2/24, 9:29 PM

@Sahil Hey! You said to ping you if needed

Ddead-brown <@319034765808369666> Why would the first message be processed by the STT? (sam...

living-lavender•6/2/24, 9:41 PM

Assistant messages aren't processed via transcriber. If you check the call logs and then search for "Transcript Output," you will notice that only users' messages are transcribed, not the assistant's response.

dead-brownOP•6/2/24, 9:41 PM

Ok so why is the transcript wrong?

living-lavender•6/2/24, 9:45 PM

Let's do one. Let's hop on a meeting and let's debug it during that process. Does that sound good to you?

dead-brownOP•6/2/24, 9:46 PM

Sure, I can be available tomorrow

living-lavender•6/2/24, 9:47 PM

Sent you appointment link.

dead-brownOP•6/2/24, 9:47 PM

What's your timezone? I'm available between 10am and 6pm in France (CEST / UTC+2)

dead-brownOP•6/2/24, 9:47 PM

Just saw the calendly in PM, thanks!

living-lavender•6/2/24, 9:48 PM

Sounds, good

instant-harlequin•12/2/25, 6:52 PM

@Sahil is this still true? i see the agent transcription happening but it's typically wrong, which then gets fed back in as model input causing errors

VapiAPP•12/19/25, 12:45 AM

Yes, it is and you can use the modelOutput if you don't want it to be transcribed.

Transcript doesn't match LLM inputs/outputs

Answer

Similar Threads

Similar Threads

Answer

Similar Threads