AfriSpeech/african-speech-id

CPU-friendly, fast language identification for 1,386 African languages # african-speech-id CPU-friendly, fast language identification for 1,386 African languages. The model uses a fast version of Omnilingual ASR to turn speech into text, and a classification head that reads that text and names the language. Both run on CPU, so a phone or a laptop is enough. The inference core is C++ with a C API, so there is no Python on the device. ```sh pip install african-speech-id ``` **How the model behaves and where it fails** is on the model card. This file is about using it. ## Results Evaluated on two held-out sets, against facebook/mms-lid-4017 on identical clips: | test set | languages | clips | **this model** | MMS-LID-4017 | |---|---|---|---|---| | Waxal | 28 | 1,675 | **0.610** | 0.604 | | omniASR corpus (test) | 76 | 4,557 | **0.429** | 0.270 | | **combined** | **104** | **6,232** | **0.478** | **0.360** | Per language the model is ahead on 57, behind on 38, and tied on 9. The gap is widest on the long tail. MMS-LID covers 4,017 languages and is strong on well-resourced ones; this model is built for the languages underneath that, which is where the omniASR test set sits. ## Quick start The head classifies text and cannot read audio, so it needs a recogniser in front of it: ```python import soundfile as sf import sherpa_onnx from african_speech_id import AfricanSpeechId model, tokens = AfricanSpeechId.download_recogniser() # omniASR, once rec = sherpa_onnx.OfflineRecognizer.from_omnilingual_asr_ctc(model=model, tokens=tokens) lid = AfricanSpeechId.load() wav, sr = sf.read("clip.wav", dtype="float32") s = rec.create_stream() s.accept_waveform(sr, wav) rec.decode_stream(s) print(lid.classify(s.result.text)) # ewe (0.03) ``` `classify()` returns `None` when no n-gram matched, meaning there was no basis for a decision. Report that as unknown rather than naming whichever language scored least badly. Labels are ISO 639-3 codes (`ewe`, `hau`, `fat`) where the source corpus carried one, and language names otherwise. …