Datasets:
Formats:
csv
Size:
< 1K
Tags:
speech-recognition
code-switching
nigerian-languages
asr-benchmark
sahara-codeswitch-challenge
License:
| import argparse | |
| import librosa | |
| import torch | |
| from transformers import AutoProcessor, Wav2Vec2ForCTC | |
| from _common import load_clips, write_results | |
| MODEL_ID = "facebook/mms-1b-all" | |
| # MMS's own ISO 639-3 language codes. | |
| MMS_LANG_CODE = { | |
| "igbo": "ibo", | |
| "yoruba": "yor", | |
| "hausa": "hau", | |
| "pidgin": "eng", # no dedicated Pidgin adapter — fall back to English, | |
| # same reasoning as Whisper's Pidgin fallback. | |
| } | |
| def main(): | |
| parser = argparse.ArgumentParser() | |
| parser.add_argument("--language", default="igbo") | |
| args = parser.parse_args() | |
| target_lang = MMS_LANG_CODE.get(args.language, "eng") | |
| if args.language == "pidgin": | |
| print("[warn] MMS has no dedicated Pidgin adapter — using 'eng'. " | |
| "Note this substitution in the report.") | |
| clips = load_clips(args.language) | |
| print(f"[info] Loading {MODEL_ID} with '{target_lang}' adapter " | |
| f"(first run downloads it — this model is large)…") | |
| processor = AutoProcessor.from_pretrained(MODEL_ID, target_lang=target_lang) | |
| model = Wav2Vec2ForCTC.from_pretrained( | |
| MODEL_ID, target_lang=target_lang, ignore_mismatched_sizes=True | |
| ) | |
| model.load_adapter(target_lang) | |
| print(f"[info] MMS: {len(clips)} clips, language '{target_lang}'") | |
| rows = [] | |
| for i, (clip_id, wav_path) in enumerate(clips, 1): | |
| audio, sr = librosa.load(wav_path, sr=16000) | |
| inputs = processor(audio, sampling_rate=16000, return_tensors="pt") | |
| with torch.no_grad(): | |
| logits = model(**inputs).logits | |
| ids = torch.argmax(logits, dim=-1)[0] | |
| transcript = processor.decode(ids).strip() | |
| rows.append((clip_id, transcript)) | |
| print(f" [{i}/{len(clips)}] {clip_id} ← \"{transcript[:60]}…\"") | |
| write_results("mms", rows) | |
| if __name__ == "__main__": | |
| main() |