adaraai/adara-speech

African speech intelligence interfaces (ASR, TTS, LID, code-switching) # adara-speech African speech intelligence interfaces: ASR, TTS, LID, diarization, code-switching, audio preprocessing. ## Status **Experimental** Classification: `INTELLIGENCE` · Visibility: `public` · License: `Apache-2.0` ## Overview Provider-agnostic interfaces for African languages and accents. Implementations may call open or commercial engines. Nothing here is a shipped production ASR. ## Why this exists Speech must be independently releasable and evaluable without pulling the whole platform. ## Quickstart ```bash pip install -e ".[dev,whisper]" # 1. Which African languages can anything installed actually handle? python scripts/coverage_report.py # 2. Make a real speech .wav using the OS's own TTS (no network, no extra deps) python scripts/make_sample_audio.py sample.wav --text "The momo no enter since yesterday" # 3. Run the whole pipeline over it python scripts/understand_audio.py sample.wav --whisper --whisper-model-size tiny # 4. Measure real backends against real African speech pip install -e ".[corpora]" # FLEURS + community corpora pip install -e ".[mms]" # MMS + community models (transformers + torch, large) python scripts/run_benchmark.py --languages sw yo ig zu --backends whisper mms --limit 8 python scripts/run_benchmark.py --languages tw --backends whisper mms community:akan-w2v2 ``` Step 4 is ADARA SpeechBench. It streams genuine recordings — FLEURS where it has the language, Ghanaian community corpora where it does not (FLEURS has no Twi) — and scores each backend independently, producing a language × backend matrix rather than a single verdict. Every result carries the corpus's exact commit SHA. Put an `HF_TOKEN` in `.env` for higher download rate limits; scripts load `.env` automatically. Step 1 is the one worth reading first. Of the 36 African languages in `src/languages/registry.py`, Whisper claims **9** (~377M speakers) and Meta's MMS claims **29**. Only a handful are reachable by neither. Those counts …