asr.py

import librosa
from transformers import Wav2Vec2ForCTC, AutoProcessor
import torch

ASR_SAMPLING_RATE = 16_000

MODEL_ID = "facebook/wav2vec2-large-960h-lv60-self"

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

def transcribe(audio):
    if audio is None:
        return "ERROR: You have to either use the microphone or upload an audio file"
    
    audio_samples = librosa.load(audio, sr=ASR_SAMPLING_RATE, mono=True)[0]
    inputs = processor(audio_samples, sampling_rate=ASR_SAMPLING_RATE, return_tensors="pt")
    
    # Set language ID for Faroese
    language_id = 'fao'  # ISO 639-3 code for Faroese
    processor.tokenizer.set_lang(language_id)

    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