Joe254h/whisper-wolof-speech

Flask web app for Wolof speech transcription using a fine-tuned Whisper model. --- title: Wolof Whisper Speech colorFrom: green colorTo: yellow sdk: docker app_port: 7860 suggested_hardware: cpu-upgrade tags: - whisper - speech-recognition - wolof - flask --- # Wolof Whisper Speech This project is a small Flask web app around a fine-tuned Whisper model for Wolof speech transcription. It keeps the original model files in `final-whisper-wolof`, includes a few WAV samples in `data`, and adds a browser interface for recording, uploading, previewing, and transcribing audio. The model was fine-tuned from `openai/whisper-small` on Wolof speech data. The app normalizes audio to the format Whisper expects before sending it to the model. ## Audio format Best input for the model: - Format: `.wav` - Also accepted by the app: `.mp3`, `.m4a`, `.flac` - Sampling frequency: 16 kHz - Audio type: mono waveform, single channel The microphone tab records in the browser, applies common capture constraints such as noise suppression and echo cancellation, then exports a 16 kHz mono WAV file before uploading it to Flask. ## Project layout - `app.py` - Flask routes and Whisper inference - `templates/index.html` - main web page - `static/css/styles.css` - app styling - `static/js/app.js` - microphone recording, upload flow, samples, and transcript actions - `final-whisper-wolof/` - local Hugging Face Transformers checkpoint - `data/` - sample WAV files and reference text - `Dockerfile` - Hugging Face Spaces deployment - `requirements.txt` - Python dependencies ## Run locally ```bash python -m venv .venv .venv\Scripts\activate python -m pip install -r requirements.txt python app.py ``` Open `localhost`. The first transcription takes longer because the model is loaded into memory on demand. After that, short clips should run faster. ## Run with Docker ```bash docker build -t wolof-whisper . docker run --rm -p 7860:7860 wolof-whisper ``` ## Hugging Face Spaces This repository is ready for a Docker Space. Hugging Face reads the YAML block at th …