Are Vapi Call recordings fully raw or do they have...
# support
a
I want to investigate conversations where the user voice wasn't captured properly i.e would like to check the difference between what the user says and what the LLM gets. My expectation would be compare recording vs transcript but want to make sure if those recordings are raw.
c
Hi, The audio is never fully raw. It is processed before delivery, as noise reduction and other audio enhancements are applied automatically.
a
and could you please be more specific with what these "enhancements" are? just to understand if e.g there might be some words spoken by the user that can get filtered out in the audio processing. what I would like to know is if these recordings fully reflect what the user actually speaks to the microphone.
also, is there a way to get the microphone blobs instead of me opening another microphone channel with .getUserMedia(), I would like to get the input volume audio. I know that right now Vapi exposes only the output audio volume (what the AI says)
can I get some help here? I just want to know if the recording I get from Vapi should accurately reflect whatever the user spoke to their microphone or how the recording is generated
c
The recordings provided by Vapi are not raw microphone audio. The audio is processed before it reaches both the recorder and the transcriber. During web calls, noise cancellation is applied on the client side, and additional processing may occur in the pipeline, such as background noise filtering, background voice filtering, endpointing detection, and interruption detection. Because of this, the recordings reflect the processed audio stream, not the exact raw microphone input. Both the transcriber and the recorder receive the same processed audio, so if you compare the recording with the transcript, any differences are typically due to speech-to-text accuracy rather than different audio sources. In some cases, very quiet speech, heavy background noise, or overlapping voices may be reduced by the processing pipeline, which could affect what ultimately gets transcribed. At the moment, Vapi does not expose raw microphone audio blobs or input audio volume levels through the SDK. If you need the raw microphone stream for comparison, the recommended approach would be to capture it separately using
getUserMedia()
alongside the Vapi call.
a
> During web calls, noise cancellation is applied on the client side, and additional processing may occur in the pipeline, such as background noise filtering can this be customized?
i've been experiencing microphone cutoffs and I believed it was because of doing this: > capture it separately using getUserMedia() alongside the Vapi call. what are other causes of a user microphone cutoff? I can provide call IDs as well
thanks for your answer btw!
c
Please share the call id and the issue you are facing