silence_cut.py

import sys
from pydub import AudioSegment, silence
import numpy as np
import matplotlib.pyplot as plt

SAMPLING_RATE = 44100
THRESHOLD = 50
CONDITION_INTERVAL = 30
OFFSET = 1

def main():
    audio_segment = AudioSegment.from_mp3(sys.argv[1])
    start_segment = audio_segment[OFFSET * 1000:15 * 1000]
    end_segment = audio_segment[-15 * 1000:-1000 * OFFSET]
    start_offset = detect_silence(start_segment.get_array_of_samples(), threshold = 50)
    end_offset = detect_silence(end_segment.get_array_of_samples(), threshold = 10, reversed = True)
    print "%lf\t%lf" % (start_offset, end_offset)

def detect_silence(segment, threshold, reversed = False):
    samples = np.array(segment)
    channel_data = samples.reshape(len(samples) / 2, 2)
    index = np.argmin(np.abs(np.average(channel_data, axis = 0)))
    sample_width = SAMPLING_RATE * 0.01
    sample_points = int(len(channel_data) / sample_width)
    means = np.zeros(sample_points)
    stds = np.zeros(sample_points)

    for i in range(sample_points):
        sample_index = i * sample_width
        data = []
        if reversed:
            data = channel_data[int(sample_index - SAMPLING_RATE * 0.05):int(sample_index), index]
        else:
            data = channel_data[int(sample_index):int(sample_index + SAMPLING_RATE * 0.05), index]
        if len(data) > 0:
            means[i] = np.average(np.abs(data))
            stds[i] = np.std(data)

    sample_count = 0
    for i in range(sample_points):
        data = []
        if reversed:
            data = (stds[-(i + CONDITION_INTERVAL):-i] - np.min(stds)) > THRESHOLD
        else:
            data = (stds[i:i + CONDITION_INTERVAL] - np.min(stds)) > THRESHOLD

        if len(data) > 0 and np.all(data):
            sample_count = i
            break
    else:
        sample_count = sample_points

    # plt.subplot(3, 1, 1)
    # plt.plot(means, color = 'g')
    # plt.subplot(3, 1, 2)
    # plt.plot((stds - np.min(stds))[0:200 * sample_width])
    # plt.subplot(3, 1, 3)
    # plt.plot(channel_data[:, index])
    # plt.show()

    return float(sample_count * sample_width) / SAMPLING_RATE + OFFSET

if __name__ == '__main__':
    main()