AfriSpeech/afrispeech-selector

Easy access to speech data across 142 African languages for training TTS and ASR models. # AfriSpeech Selector **AfriSpeech Selector gives you access to 142 languages · 2267.9 hours · 35 countries of African speech for training TTS and ASR models.** Select languages by recorded **hours** (strength) — a country-balanced top-N or a hand-picked set, sized the way you want — and get the **audio + metadata in the format your training pipeline expects**. You take it from there; the tool doesn't do any text normalisation or cleaning (that's your framework's job). It's clean read speech with aligned transcripts, so the natural fit is **TTS** — export WAVs + a manifest for **LJSpeech, Piper, VITS, or MeloTTS**. It works for **ASR** too (`load_from_disk` / Parquet, or stream with `stream_dataset(...)`), which is handy for supplementing low-resource languages. **Redistribution:** building a *local* working set for your own training is fine. Redistributing copies of the audio (e.g. `--push` to a public repo) is **not recommended**, given the permissions of the underlying public data sources. ## Available languages **142 languages · 2267.9 hours · 35 countries.** Hours is the strength signal used for ranking. The `--languages` value is the name you pass to pick a language (e.g. `--languages twi_twi`). List them anytime with `afrispeech-select --list-langs`, or see the full **language catalog** at the bottom. ## Install ```bash git clone github.com cd afrispeech-selector python3 -m venv .venv && source .venv/bin/activate pip install -e . # gives you the `afrispeech-select` command ``` (Use `python3` — the code uses non-ASCII text and won't run under Python 2.) ## Quickstart — one language, ready to train Most people just want one language for **TTS**. Pick its name (`afrispeech-select --list-langs`, or the catalog at the bottom) and run one command. Clips are filtered to a **3–15 s** window by default, so the result is training-ready. ```bash # ~5 hours of Twi as an LJSpeech TTS dataset (wavs/ + metad …