Timeout errors with whisper-large and long processing times.

[nishanthcmesh]

While using whisper-cloud with the whisper-large model, I kept getting timeout errors for a ~1 hr audio file.
I even tried breaking the file into chunks of 2.5 minutes and the errors continued, for ~7 chunks out of 23 that I had.
In addition, latency was very high, taking 48 minutes for a 57 minute file, with some segments timing out.
The chunks were audio/mp4 files of 2.5 minutes each for context.

1. Is there some kind of fix for these timeouts and high latencies?

2. How do I set the timeout parameter using the deepgram python sdk?

3. Is there some kind of rate limit/higher latency for free tier accounts? We are considering using Deepgram in our app and it is unusable with these many errors and timeouts. I checked the status page for both times when I tried and DG whisper cloud was listed as operational.

4. Whisper-large transcription quality was way better for us compared to Nova. We were trying with Indian English speakers, is Nova not optimized for Indian accents?

[uanandaraja]

The latency and timeout error for Whisper has been occurring for me too for the last 4 days, at least. Hopefully, this get resolved soon. Maybe no one is resolving it due to weekend?

[jkroll-deepgram]

Hi @nishanthcmesh and @uanandaraja, intermittent higher Whisper latency is a known occurrence as we manage our overall traffic. Your approach to chunk files will help, as will converting any video files to audio-only (if applicable). Nova will have much lower latency than Whisper.

Nova supports several dialects of English. For Indian English speakers, using en-IN rather than en or en-US will increase accuracy. See our full Languages docs.