asr.py

import torch
import torchaudio
import torchaudio.functional as AF
from transformers import Wav2Vec2ForCTC, AutoProcessor
from pydub import AudioSegment
from pydub.silence import split_on_silence


class Transcribe:
    def __init__(self, freq: float = 16000.0) -> None:
        self.freq = freq
        self.model_id = "facebook/mms-1b-fl102"
        self.processor = AutoProcessor.from_pretrained(self.model_id)
        self.model = Wav2Vec2ForCTC.from_pretrained(self.model_id)

    @torch.inference_mode()
    def __call__(self, audio_tensor: torch.tensor, lang: str = "amh"):
        print(lang)
        self.processor.tokenizer.set_target_lang(lang)
        self.model.load_adapter(lang)

        outputs = self.model(audio_tensor)
        logits = outputs.logits
        ids = torch.argmax(logits, dim=-1)[0]
        decoded_token = self.processor.decode(ids)

        return decoded_token