|
|
--- |
|
|
language: ml |
|
|
library_name: mlx |
|
|
license: apache-2.0 |
|
|
pipeline_tag: automatic-speech-recognition |
|
|
tags: |
|
|
- whisper |
|
|
- mlx |
|
|
- Automatic Speech Recognition |
|
|
base_model: vrclc/Whisper-medium-Malayalam |
|
|
--- |
|
|
|
|
|
## Whisper-medium-Malayalam (MLX) |
|
|
|
|
|
Apple MLX-converted weights for `vrclc/Whisper-medium-Malayalam` optimized for Apple Silicon. |
|
|
|
|
|
- Base model: `vrclc/Whisper-medium-Malayalam` |
|
|
- Format: MLX (`weights.safetensors`, `config.json`) |
|
|
- Intended runtime: `mlx-whisper` on Apple Silicon (M-series) |
|
|
|
|
|
### Usage (Python) |
|
|
```python |
|
|
import mlx_whisper |
|
|
|
|
|
result = mlx_whisper.transcribe( |
|
|
"/path/to/audio.wav", |
|
|
path_or_hf_repo="<this-repo>", |
|
|
# Optional decoding controls |
|
|
language="ml", # Malayalam |
|
|
task="transcribe", # or "translate" |
|
|
temperature=0.0, |
|
|
no_speech_threshold=0.3, |
|
|
logprob_threshold=-1.0, |
|
|
compression_ratio_threshold=2.4, |
|
|
) |
|
|
print(result["text"]) |
|
|
``` |
|
|
|
|
|
### Local HTTP server (FastAPI) |
|
|
With the server at `whisper/server_mlx.py` from `avatar-npm`: |
|
|
```bash |
|
|
export WHISPER_MODEL=<this-repo-or-local-mlx-path> |
|
|
export WHISPER_LANGUAGE=ml |
|
|
python server_mlx.py |
|
|
# POST /transcribe with form field `file` |
|
|
``` |
|
|
|
|
|
### Notes |
|
|
- This repo contains only the MLX weights and config. Tokenization and audio |
|
|
preprocessing are handled by `mlx-whisper`. |
|
|
- If you need the original (non-MLX) model, see `vrclc/Whisper-medium-Malayalam`. |
|
|
|
|
|
### License |
|
|
The original model’s license applies. See the upstream repository for details. |
|
|
|
|
|
|
|
|
|
