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(audio_source=None, microphone=None, file_upload=None):
    audio_fp = file_upload if file_upload else microphone
    if audio_fp is None:
        return "ERROR: You have to either use the microphone or upload an audio file"
    
    audio_samples = librosa.load(audio_fp, sr=ASR_SAMPLING_RATE, mono=True)[0]

    # Set Faroese language
    processor.tokenizer.set_target_lang("fao")
    model.load_adapter("fao")

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

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device)
    inputs = inputs.to(device)

    with torch.no_grad():
        outputs = model(**inputs).logits
        ids = torch.argmax(outputs, dim=-1)[0]
        transcription = processor.decode(ids)

    return transcription