aaderemi/yor_s2tt

Yoruba speech to english text speech to text translation task. # Yoruba → English Speech Translation Fine-tunes OpenAI Whisper (`small`) to translate spoken Yoruba audio into English text. The encoder is updated while the decoder is frozen, making this an efficient transfer-learning setup. --- ## Requirements ```bash pip install torch transformers datasets librosa numpy pandas tqdm nltk huggingface_hub ``` A CUDA-capable GPU is strongly recommended. The script automatically falls back to CPU but training will be very slow without a GPU. --- ## Data Format ### Audio Pre-extracted audio arrays stored as `.npy` files (float32, mono, 16 kHz). Expected directory layout: ``` yoruba/ train/audio/ ← .npy files, e.g. Y_001_0001.npy dev/audio/ ← .npy files ``` File naming convention: `Y_ _ .npy` The script maps each Yoruba file to its English counterpart using the ID scheme `E _ `. ### Labels (CSV) Two CSV files with at minimum these two columns: | column | description | |----------|-------------------------------------| | `text_id`| Unique utterance ID, e.g. `E01_0001`| | `text` | English translation string | ``` english_text/ train/eng_text_train.csv dev/eng_text_dev.csv ``` --- ## Usage ### Basic (all defaults) ```bash python train.py ``` ### Custom paths and hyperparameters ```bash python train.py \ --model_id openai/whisper-medium.en \ --train_audio_dir yoruba/train/audio/ \ --train_csv english_text/train/eng_text_train.csv \ --dev_audio_dir yoruba/dev/audio/ \ --dev_csv english_text/dev/eng_text_dev.csv \ --output_dir s2tt_models \ --epochs 5 \ --train_batch_size 16 \ --eval_batch_size 16 \ --learning_rate 1e-3 \ --warmup_steps 500 \ --weight_decay 0.001 ``` ### All CLI arguments | Argument | Default | Description | |---------------------|--------------------------------------|------------------------------------------| | `--model_id` | `openai/whispe …