app.py

import torch
import torchaudio
from fastapi import FastAPI, UploadFile, File
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
import io

app = FastAPI()

# Load Wav2Vec2 model and processor
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-lv-60-espeak-cv-ft")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-lv-60-espeak-cv-ft")


@app.post("/transcribe/")
async def transcribe_audio(file: UploadFile = File(...)):
    try:
        # Load audio file
        audio_bytes = await file.read()
        audio_input, sample_rate = torchaudio.load(io.BytesIO(audio_bytes))

        # Convert stereo to mono (if needed)
        if audio_input.shape[0] > 1:
            audio_input = torch.mean(audio_input, dim=0, keepdim=True)

        # Resample to 16 kHz (if needed)
        target_sample_rate = 16000
        if sample_rate != target_sample_rate:
            resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_sample_rate)
            audio_input = resampler(audio_input)

        # Remove batch dimension
        audio_input = audio_input.squeeze(0)

        # Preprocess the audio
        input_values = processor(audio_input, sampling_rate=target_sample_rate, return_tensors="pt").input_values

        # Run inference
        with torch.no_grad():
            logits = model(input_values).logits

        # Decode the predicted tokens
        predicted_ids = torch.argmax(logits, dim=-1)
        transcription = processor.batch_decode(predicted_ids)

        return {"transcription": transcription[0]}

    except Exception as e:
        return {"error": str(e)}