go-whisper golang bindings

[djthorpe]

Create bindings for https://github.com/ggerganov/whisper.cpp

• [x] Simple golang bindings with tests
• [x] Some examples (main, sample) based off of these
• [x] Integrate with ffmpeg for audio conversion
• [x] Some sort of real-time translation
• [x] gRPC and/or websocket API
• [x] Docker image of a speech-to-tech service

[djthorpe]

Made PR: https://github.com/ggerganov/whisper.cpp/issues/269

[chrisbward]

Great work!

Keen on realtime translation and a way of calling out/streaming the output to another app - gRPC seems the best option for this

[djthorpe]

Yeah thanks.

I'm doing the audio downsampling to 16KHz at the moment in a different repository (go-media)

The realtime transcription and translation should be pretty straightforward, but pretty experimental, even for whisper.cpp

I will take a while to get to the gPRC microservice :-(

[djthorpe]

Added a "stream" command for the start of real-time streaming, but:

• Thread safety: Needs some work to ensure the same model can be used in the process method across threads/goroutines
• Ring buffer: Implement a ring buffer for continious audio samples
• Overlaps: Need some word overlaps to ensure we don't lose words between sample windows
• Silence: Don't process audio when silence is fed in. Ideally chunk windows when there is a largish (>1s) silence

There's also some issues with the segmenting in the main package (repeated segments come out!) needs fixing.

[djthorpe]

Coming back to this after some time!

Remaining tasks:

• [x] Streaming output of segments from the server
• [ ] Streaming input of audio from the client and an example of this working
• [ ] Add iniital token prompts for segments
• [x] Have ffmpeg logging go through the whisper logging (single source for all log messages)
• [ ] Not sure MaxConcurrent tasks is really working - need to check
• [x] Fix Dockerfiles so they work
• [x] Add Diarization
• [x] SRT/VTT and text output as well as JSON
• [ ] Add tokens into the verbose_json format
• [x] Fix the README so it reflects reality
• [x] Investigate the segmenter having bad output (seems to work, but the language detection makes it look off)

Lower priority:

• [ ] Integrate VAD voice detection (probably from https://github.com/baabaaox/go-webrtcvad which seems to work) on segments
• [ ] Smarter segmentation so that it only happens on silence boundaries, not in the middle of words
• [ ] Fix model id's so they are unique and that models can be stored at some sub-path
• [ ] Integrate the bindings back into whisper.cpp repository

[djthorpe]

Also:

• [ ] Fix client so that it works with text/event-stream for both downloading models and transcription
• [x] Logging is generally not working. Fix it so only messages apart from errors are supressed unless debug mode. Also output logging of requests
• [ ] Add some metrics in there somewhere

[djthorpe]

Also:

• [x] Fix resampling of raw audio in the go-media code, so we can again ingest WAV files without "input changed" errors

[djthorpe]

Simplified Dockerfile and now uses the base images from here as a base:

https://github.com/mutablelogic/docker-llamacpp

This is still now working; Now I need to have the ffmpeg shared libraries included in the runtime image. Considering whether to just copy over the libraries from the build image, or to install ffmpeg libraries from source.