utils/audio_preprocessor.py

import numpy as np
import librosa

def advanced_preprocess(audio_data, sr):
    """
    Applies the same preprocessing used in the advanced training pipeline:
    1. Spectral Gating Noise Reduction
    2. Pre-emphasis filter
    3. Normalization
    """
    # 1. Noise Reduction
    D = librosa.stft(audio_data)
    magnitude, phase = np.abs(D), np.angle(D)
    noise_floor = np.percentile(magnitude, 10, axis=1, keepdims=True)
    mask = np.maximum(magnitude - 1.5 * noise_floor, 0) / (magnitude + 1e-8)
    D_cleaned = magnitude * mask * np.exp(1j * phase)
    audio_clean = librosa.istft(D_cleaned, length=len(audio_data))
    
    # 2. Pre-emphasis
    audio_pre = np.append(audio_clean[0], audio_clean[1:] - 0.97 * audio_clean[:-1])
    
    # 3. Normalization
    max_val = np.max(np.abs(audio_pre))
    if max_val > 0:
        return audio_pre / max_val
    return audio_pre