utils.py

import torch
import torchaudio

def load_model(model_path, model_class, num_classes=10):
    model = model_class(num_classes=num_classes)
    model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
    model.eval()
    return model

def preprocess_audio(file_path):
    waveform, sample_rate = torchaudio.load(file_path)
    transform = torchaudio.transforms.MelSpectrogram(
        sample_rate=sample_rate,
        n_mels=64,
        n_fft=1024,
        hop_length=512
    )
    mel_spec = transform(waveform)
    mel_spec = mel_spec.mean(dim=0, keepdim=True)
    return mel_spec

def predict(model, input_tensor, labels):
    with torch.no_grad():
        outputs = model(input_tensor.unsqueeze(0))
        _, predicted = torch.max(outputs.data, 1)
    return labels[predicted.item()]