Adom-Amponsah/Ghana-tts

building our tts # Ghana-tts Quality audit of the Hugging Face dataset `ghanaopendata/ghana-english-tts-filtered` (~303,204 clips / ~1,142 h of Ghanaian news-broadcast speech, 16 kHz mono) to decide whether it is training-grade material for a Ghanaian-English TTS voice. **Verdict: GO.** With a `duration_ss >= 12 s` filter the corpus yields an estimated **~925-955 h of clean speech** (~85-88% keep rate, verified by two independent human listening audits). ## What's here | Path | Content | |---|---| | `scripts/` | every step, reproducible: dataset discovery, stratified sampling, DSP QC, listening-sheet builders, duration profiling, verdict compilers | | `reports/` | `dataset_overview.md`, `sample_quality_report.md` (full audit results), audit workbooks, listening sheets | | `data/manifests/` | per-shard footer stats, sample manifests, full-corpus duration profile (`.csv.gz`) | | `analysis/` | automated DSP QC features for the 200-clip sample | Audio samples are **not** included (gitignored); they are re-extractable from the HF dataset via `scripts/build_sample.py` / `scripts/random_pass_12s.py`. ## Key findings - The HF viewer fails on this dataset: shard names produce the illegal split name `filtered-train`. All access goes through direct parquet reads (`hf://` URLs) - see `scripts/dataset_info.py`. - Filter fingerprints: hard 15.000 s ceiling, 0.85 speech-prob floor, -30 dBFS floor. Only 0.9% of clips sit exactly at the ceiling. - Clip duration is the dominant quality predictor: 12% keep below 4 s vs 92% keep at >= 12 s (pass 1, n=200). Pass 2 on random shards confirmed 81% keep + 13% borderline in the 12-15 s band. - DSP energy heuristics are poor predictors of human judgment; use the duration floor instead. ## Environment Python 3.13 venv (`.venv/`, gitignored). Deps: `datasets`, `huggingface_hub`, `pandas`, `pyarrow`, `soundfile`, `librosa`, `numpy`, `matplotlib`, `tqdm`, `openpyxl`.