src/utils/audio_helper.py

import io
from fastrtc.utils import audio_to_int16
from pydub import AudioSegment
import wave
import torch
import torchaudio
import numpy as np

def audio_to_bytes(audio_tuple, sample_rate=24000) -> io.BufferedReader:
        sr, audio_data = audio_tuple
        audio_int16 = audio_to_int16(audio_tuple)

        buffer = io.BytesIO()
        with wave.open(buffer, "wb") as wf:
            wf.setnchannels(1)
            wf.setsampwidth(2)
            wf.setframerate(sr)
            wf.writeframes(audio_int16.tobytes())
        buffer.seek(0)
        buffer.name = "audio.wav"
        return buffer

def resample_audio(audio_buffer:io.BytesIO):
    audio_buffer.seek(0)

    audio_segment = AudioSegment.from_file(audio_buffer, format="mp3")
    samples = np.array(audio_segment.get_array_of_samples()).astype(np.float32) / (2 ** 15)
    
    
    if audio_segment.channels == 2:
        samples = samples.reshape((-1, 2)).mean(axis=1)
    
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    
    audio_tensor = torch.from_numpy(samples).unsqueeze(0).to(device)  

    resampler = torchaudio.transforms.Resample(
        orig_freq=audio_segment.frame_rate,
        new_freq=24000
    ).to(device)
    
    resampled_tensor = resampler(audio_tensor)
    resampled = resampled_tensor.squeeze(0).cpu().numpy()

    return resampled