asr.py

import librosa
from transformers import Wav2Vec2ForCTC, AutoProcessor
import torch

ASR_SAMPLING_RATE = 16_000


MODEL_ID = "facebook/mms-1b-all"

processor = AutoProcessor.from_pretrained(MODEL_ID)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)


def transcribe(microphone, file_upload, lang):

    warn_output = ""
    if (microphone is not None) and (file_upload is not None):
        warn_output = (
            "WARNING: You've uploaded an audio file and used the microphone. "
            "The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
        )
    elif (microphone is None) and (file_upload is None):
        return "ERROR: You have to either use the microphone or upload an audio file"

    audio_fp = microphone if microphone is not None else file_upload
    audio_samples = librosa.load(audio_fp, sr=ASR_SAMPLING_RATE, mono=True)[0]

    lang_code = lang.split(":")[0]
    processor.tokenizer.set_target_lang(lang_code)
    model.load_adapter(lang_code)

    inputs = processor(
        audio_samples, sampling_rate=ASR_SAMPLING_RATE, return_tensors="pt"
    )

    with torch.no_grad():
        outputs = model(**inputs).logits

    ids = torch.argmax(outputs, dim=-1)[0]
    transcription = processor.decode(ids)
    return warn_output + transcription