src/speech_to_text.py

import librosa
import torch
from transformers import Wav2Vec2ForCTC, AutoProcessor
from transformers import set_seed


def transcribe(fp:str, target_lang:str) -> str:
    ''' 
    For given audio file, transcribe it.
    
    Parameters
    ---------- 
    fp: str
        The file path to the audio file.
    target_lang:str
        The ISO-3 code of the target language.

    Returns
    ---------- 
    transcript:str
        The transcribed text.
    '''
    # Ensure replicability
    set_seed(555) 

    # Load transcription model
    model_id = "facebook/mms-1b-all"
    target_lang = "mos"

    processor = AutoProcessor.from_pretrained(model_id, target_lang=target_lang)
    model = Wav2Vec2ForCTC.from_pretrained(model_id, target_lang=target_lang, ignore_mismatched_sizes=True)

    # Process the audio
    signal, sampling_rate =  librosa.load(fp, sr=16000)
    inputs = processor(signal, sampling_rate=16_000, return_tensors="pt")
    
    # Inference
    with torch.no_grad():
        outputs = model(**inputs).logits
    
    ids = torch.argmax(outputs, dim=-1)[0]
    transcript = processor.decode(ids)

    return transcript