noise_reducer.py

import numpy as np
import soundfile as sf
import noisereduce as nr
from scipy.signal import butter, filtfilt


def butter_bandpass(lowcut=80, highcut=8000, fs=16000, order=5):
    """Keep only human voice frequencies (80Hz–8kHz)."""
    nyq = fs / 2
    low = lowcut / nyq
    high = min(highcut / nyq, 0.99)
    b, a = butter(order, [low, high], btype='band')
    return b, a


def reduce_noise(
    input_wav: str,
    output_wav: str = "cleaned_audio.wav",
    prop_decrease: float = 0.75,
    use_bandpass: bool = True,
    normalize: bool = True
) -> str:
    """
    Apply configurable audio preprocessing pipeline.

    Args:
        input_wav:      Input WAV file path
        output_wav:     Output WAV file path
        prop_decrease:  Noise reduction aggressiveness (0.0–1.0). 0.0 = skip noise reduction.
        use_bandpass:   Apply bandpass filter (80Hz–8kHz voice range)
        normalize:      Normalize audio volume
    """
    data, rate = sf.read(input_wav)

    # Step 1: Bandpass filter — keep human voice frequencies only
    if use_bandpass:
        try:
            b, a = butter_bandpass(fs=rate)
            data = filtfilt(b, a, data)
        except Exception:
            pass  # skip if filter fails on short audio

    # Step 2: Noise reduction
    if prop_decrease > 0.0:
        noise_sample_duration = int(0.5 * rate)
        noise_clip = data[:max(noise_sample_duration, 1)]
        data = nr.reduce_noise(
            y=data,
            sr=rate,
            y_noise=noise_clip,
            prop_decrease=prop_decrease,
            stationary=False,
            n_fft=2048,
            n_std_thresh_stationary=1.5
        )

    # Step 3: Normalize volume
    if normalize:
        max_val = np.max(np.abs(data))
        if max_val > 0:
            data = data / max_val * 0.95

    sf.write(output_wav, data, rate)
    return output_wav