GhanaNLP/ghana-ipa-asr

Speech to IPA phonemes for 42 Ghanaian and West African languages — CTC model, CPU/GPU/mobile inference via sherpa-onnx # ghana-ipa-asr Speech to **IPA phonemes** for **42 Ghanaian and West African languages**, in one model. ```bash pip install git+github.com ghana-ipa-asr utterance.wav ``` ``` m f ia s e n o o ɲ a n kʰ o pʰ ɔ n b ɔ ɔ ɔ s o ɾ o . ``` | | | |---|---| | Languages | 42, one shared model | | Output | IPA phoneme stream, with punctuation | | Accuracy | 16.64% UER held-out (15.39% excluding punctuation) | | Size | **349 MB** int8, 1.3 GB fp32 | | Runs on | CPU, NVIDIA GPU, Apple CoreML, Android/iOS | | Corpus scale | `ghana-ipa-asr batch` — **~2100× realtime** on one H200 | It is a CTC model, so its alignment is **monotonic — it cannot emit phonemes that are not in the audio**. For phonetic work that guarantee matters more than raw accuracy: an invented symbol is worse than a missing one. ## Output formats The model emits a flat list of units; every format is a view over that list, so switching costs nothing and never re-runs the model. ```python from ghana_ipa_asr import GhanaIPAASR asr = GhanaIPAASR.load() t = asr.transcribe("utterance.wav") t.spaced() # 'm f ia s e n o o ɲ a n kʰ o pʰ ɔ n' every unit separated t.continuous() # 'mfiasenooɲankʰopʰɔnbɔɔɔsoɾo' no separators at all t.units() # ['m', 'f', 'ia', 's', 'e', ...] as a list t.spaced(punctuation=False) # drop . , ? ! and the rest ``` From the command line: ```bash ghana-ipa-asr audio.wav --format spaced # default; also what forced alignment needs ghana-ipa-asr audio.wav --format continuous # bare phoneme string ghana-ipa-asr audio.wav --format json # every format at once ghana-ipa-asr *.wav --no-punctuation ``` **Which format to use** | you want | use | why | |---|---|---| | forced alignment, per-phoneme timings | `spaced()` | units like `kʰ` and `k͡p` are more than one character, so a character-level split would turn one sound into two | | a bare phoneme string | `continuous()` | no separators; note un …