Use VAD to separate out regions without speech content, and estimate mean power of background noise

speech_noise.py

"""
Estimate background noise power level of a speech waveform.
Requires some non-speech regions in the wave.

Requirements:
    pip install numpy librosa soundfile webrtcvad

MIT License John Meade 2021
"""


import librosa
import numpy as np
import soundfile as sf
import webrtcvad
import struct


int16_max = (2 ** 15) - 1


def float_to_pcm16(wav):
    "pack a wav, in [-1, 1] numpy float array format, into 16 bit PCM bytes"
    return struct.pack("%dh" % len(wav), *(np.round(wav * int16_max)).astype(np.int16))


def nonspeech(wav, sr, vad_aggressiveness=3, webrtc_sr=16_000, webrtc_window_ms=30, dilations=3):
    """
    Args:
        wav: the waveform, a float numpy array with range [-1, 1].
        sr: sample rate of input wav.
        vad_aggressiveness: must be 1, 2 or 3. Higher is more strict about filtering nonspeech.
        webrtc_sr: must be 8000, 16000, 32000, or 48000. Use a value close to your input.
        webrtc_window_ms: must be 10, 20, or 30ms.
        dilations: number of windows to dilate VAD results. Increase if phonemes are leaking into output.
    Returns:
        wave of all regions with no speech content, concatenated together
    """

    # resample
    if sr != webrtc_sr:
        wav = librosa.core.resample(wav, sr, webrtc_sr)

    # init VAD
    vad = webrtcvad.Vad(vad_aggressiveness)

    # trim wav to integer number of windows
    W = webrtc_window_ms * webrtc_sr // 1000
    T = len(wav)
    rem = T % W
    wav = wav if rem == 0 else wav[:-rem]

    # run VAD
    windows = [wav[i * W:(i+1) * W] for i in range(len(wav) // W)]
    va = [vad.is_speech(float_to_pcm16(win), webrtc_sr) for win in windows]

    # dilate VAD to adjacent frames
    va = np.array(va, dtype=bool)
    for _ in range(dilations):
        va[:-1] |= va[1:]
        va[1:] |= va[:-1]

    # collect all frames without VA, concat into background-audio wave
    bg = [win for win, is_speech in zip(windows, va) if not is_speech]
    bg = bg if len(bg) == 0 else np.concatenate(bg)
    return bg


def mean_noise_power(wav_fpath):
    # read wav
    wav, sr = sf.read(wav_fpath)

    # normalize
    wav /= abs(wav).max().clip(1e-6)

    # obtain nonspeech samples
    bg = nonspeech(wav, sr)

    # compute mean power
    return (bg ** 2).mean()


if __name__ == '__main__':
    # Example, write nonspeech to file to verify that the VAD performed correctly
    wav, sr = sf.read('path/to/file.wav')
    wav /= abs(wav).max().clip(1e-6)
    bg = nonspeech(wav, sr, webrtc_sr=16_000)
    sf.write('path/to/file_nonspeech.wav', bg, 16_000)

These (this and WADA SNR) are very useful SNR scripts. Do you have a usage license for them?

Apologies just seen the license at the top of the code.