LateefAdeleke/Yoruba_Transcription

# Yoruba_Transcription # Yoruba Checkpoint Transcription #This is the minimal setup for transcribing Yoruba audio with Lateef's best Whisper checkpoint (This checkpoint achieves 0.3 CER): `/workspace/storage/results/yoruba_pretrain/checkpoint-3520` ## 1) Install dependencies ```bash pip install torch transformers librosa soundfile sentencepiece numpy ``` If audio decoding is incomplete on your machine, also install `ffmpeg`. ## 2) Transcribe one audio file ```bash python scripts/transcribe_yoruba_checkpoint.py /path/to/audio.wav ``` ## 3) Transcribe a folder of audio files ```bash python scripts/transcribe_yoruba_checkpoint.py /path/to/audio_folder --output-csv yoruba_transcriptions.csv ``` ## 4) Use a different checkpoint path ```bash python scripts/transcribe_yoruba_checkpoint.py /path/to/audio.wav --checkpoint /path/to/checkpoint ``` ## Notes - The script forces Yoruba transcription with `language="yo"` and `task="transcribe"`. - Supported audio extensions: `.wav`, `.mp3`, `.flac`, `.m4a`, `.ogg`, `.opus` - For a folder input, the script prints each transcription and can also save a CSV with: - `audio_path` - `transcription`