Whisper V3 model can support long Audio input, can you add an API to support large Audio file as a whole?

[ER-EPR]

Is your feature request related to a problem? Please describe.

Current whisper api can only take around 20MB audio file per request. Now whisper V3 can work with large audio file. Can you support it too. Eliminate the limitation on audio file size for whisper V3.
Describe the solution you'd like

Allow upload a large audio file.
call whisper with V3 model, transcribe.
Stream the text back to the user.
Describe alternatives you've considered

Additional context

[mudler]

good point, we need to update whisper

[ER-EPR]

good point, we need to update whisper

but I still can't send 17MByte MP3 audio file to the API :

ClientException
Client error: `POST http://192.168.110.70:8080/v1/audio/transcriptions` resulted in a `413 Request Entity Too Large` response: {"error":{"code":413,"message":"Request Entity Too Large","type":""}}
API request error : Client error: `POST http://192.168.110.70:8080/v1/audio/transcriptions` resulted in a `413 Request Entity Too Large` response: {"error":{"code":413,"message":"Request Entity Too Large","type":""}}

I'm using ggml-whisper-largev3 model with localai:v2.11.0-cublas-cuda12-ffmpeg
Is there something else needs to be changed? @mudler

[sfxworks]

curl http://172.16.1.193/v1/audio/transcriptions -H "Content-Type: multipart/form-data" -F file="@/tmp/audio.mp3" -F model="whisper-3"
{"error":{"code":413,"message":"Request Entity Too Large","type":""}}%

Yeah even a direct curl on a v3 model. I'm using the all in one cuda 12 image here.

[sfxworks]

Downsampled audio file and saved to /tmp/audio_downsampled.mp3 with bitrate 32k
Audio file size in MB: 16.477432

Even under that size,

openai.APIStatusError: Error code: 413 - {'error': {'code': 413, 'message': 'Request Entity Too Large', 'type': ''}}

[sfxworks]

Hey so its currently undocumented but found it in discord

UPLOAD_LIMIT needs to be set to something higher. Set it to 50 and it works.

[ER-EPR]

set environment LOCALAI_UPLOAD_LIMIT right? Can I set it to 512?

[sfxworks]

donno but I've managed to upload at least to 100 give it a shot

[fullstackwebdev]

just for future reference, yes it works

docker run -p 8080:8080 --gpus all --name local-ai2 -ti -e LOCALAI_UPLOAD_LIMIT=512 localai/localai:latest-aio-gpu-nvidia-cuda-12