blockvoltcr7/darija-model-speech-to-text

Serverless Moroccan Darija speech-to-text API on Modal # Darija ASR on Modal Deploy a serverless Moroccan Darija speech-to-text API on Modal using a Whisper Large v3 Turbo base model and a Darija LoRA adapter. The deployed service exposes: ```text POST /transcribe ``` Request body: ```json { "filename": "sample.m4a", "audio_base64": "..." } ``` Response body: ```json { "text": "دارجة مكتوبة بالعربية", "model": "anaszil/whisper-large-v3-turbo-darija", "base_model": "openai/whisper-large-v3-turbo", "filename": "sample.m4a", "audio_bytes": 123456, "duration_seconds": 2.345 } ``` ## Model - Base model: `openai/whisper-large-v3-turbo` - Darija LoRA adapter: `anaszil/whisper-large-v3-turbo-darija` - Runtime: Modal serverless GPU, default `L4` - Supported audio formats: WAV, MP3, M4A, OGG, and other formats supported by `ffmpeg` The Darija adapter model card lists the adapter as MIT licensed and reports WER around 24.88% and CER around 8.28% on its evaluation set. Review the upstream model cards before using this in production. ## Prerequisites Install: - Python `3.11+` - `uv` - A Modal account Install `uv` if needed: ```bash curl -LsSf astral.sh | sh ``` Authenticate Modal: ```bash uv run modal setup ``` ## Setup Clone the repository and install local tooling: ```bash git clone cd darija-model-speech-to-text uv sync ``` Create your local environment file: ```bash cp .env.example .env ``` Generate a test API key and put it in `.env`: ```bash python - --transcribe.modal.run ``` Check the deployed service: ```bash curl https:// --transcribe.modal.run/healthz ``` Expected response: ```json { "status": "ok", "model": "anaszil/whisper-large-v3-turbo-darija", "base_model": "openai/whisper-large-v3-turbo" } ``` ## Test With Audio Put local test audio files in: ```text entry-point-audio-files/ ``` This directory is ignored by git so private audio is not committed. Run a transcription test: ```bash uv run python scripts/test_endpoint.py \ --url "https:// --transcribe.modal.run" \ …