dogahwisdom/ghana-speech-asr

Production Whisper ASR fine-tuning for Ghana Speech (Asante/Akuapem Twi, Fante, Ewe) — Wisdom Dogah # Ghana Speech ASR — Wisdom Dogah Production-grade Whisper fine-tuning pipeline for the Ghana Speech dataset on NVIDIA H200 (Ghana NLP / AISCA). **Author:** Wisdom Dogah **License (models / data derivatives):** CC BY-NC 4.0 **Target languages (v1):** Asante Twi · Akuapem Twi · Fante · Ewe This repository is intentionally structured like a serious ML training codebase — config-driven experiments, indexed manifests, reproducible splits, WER/CER evaluation, and Hub publishing — not a one-off notebook. ## Why this design | Practice | Implementation | |---|---| | Config-driven runs | `configs/*.yaml` | | Leakage-aware splits | Hash-bucket by `language + source_file` | | Indexed I/O | Manifest stores `parquet_path` + `row_index` | | Robust audio decode | `soundfile` on embedded bytes (no `torchcodec` required) | | Standard trainer | Hugging Face `Seq2SeqTrainer` + Whisper | | Metrics that matter | WER + CER via `evaluate` / `jiwer` | | Session safety | Checkpoints + TensorBoard under `outputs/`; push to HF before wipe | ## Ghana NLP session rules (followed) From the H200 quickstart / model-card template: 1. Dataset lives at `/data/ghana-speech` (42 language subsets, 16 kHz mono). 2. License is **CC BY-NC 4.0** — non-commercial only. 3. Share trained models with **`ghananlpcommunity`** on Hugging Face and include a model card. 4. `/workspace` is **wiped** at window end — export / push before the deadline. 5. Use **`tmux`** for long-running jobs. ## Repository layout ```text configs/ Experiment YAML src/ghana_asr/ cli/ prepare | train | evaluate | push data/ manifests + indexed dataset + collator evaluation/ WER/CER training/ Seq2Seq training loop utils/ seeding, audio, logging model_card/MODEL_CARD.md Hub README template (Ghana NLP format) scripts/ Thin wrappers for H200 / tmux tests/ Unit tests for split hashing & config ``` ## Q …