audio_preprocess.py

'''
This script takes a wave file, detects and removes any leading silence, normalizes it then slices it into a 30 second segment and returns that segment.
Code based on the answers found here:
https://stackoverflow.com/questions/29547218/remove-silence-at-the-beginning-and-at-the-end-of-wave-files-with-pydub
https://stackoverflow.com/questions/37999150/how-to-split-a-wav-file-into-multiple-wav-files
The functions that are commented out would be useful if the audio has different sample rates and are mono+stereo
''' 
from pydub import AudioSegment, effects
from scipy.io import wavfile
import torch
import torchaudio
from torchaudio import transforms


def audio_preprocess(audio_file,trim_start,trim_end):

    # Path to the wav file track, can be adjusted
    sound = AudioSegment.from_file(audio_file, format="wav")
    # Defines the start and end of the silence
    start_trim = leading_silence(sound)
    # Defines the length of the silence
    duration = len(sound)    
    # Resulting sound file 
    trimmed_sound = sound[start_trim:duration]
    # Normalize the trimmed audio file
    normalized_sound = effects.normalize(trimmed_sound)
    '''
    # Resample the sound to the desired rate
    resampled_sound = resample(normalized_sound,44100)
    # Rechannel to stereo
    stereo_sound = rechannel(resampled_sound, 2)
    # The new audio file trimmed from defined start and end
    '''
    new_audio = normalized_sound[trim_start:trim_end]
    return new_audio

# Function to detect the leading silence of a track
def leading_silence(sound, silence_threshold=-50.0, chunk_size=10):
    # Temp variable stores length of the trim
    trim_ms = 0 
    # Avoid infinite loop
    assert chunk_size > 0 
    # While the track is below the since threshold, start trimming the sound length 
    while sound[trim_ms:trim_ms+chunk_size].dBFS < silence_threshold and trim_ms < len(sound):
        trim_ms += chunk_size

    return trim_ms

'''
# Resample the first then second channel and merges both channels
# Function returns sound if in desired sample rate
# Outputs a tensor
def resample(sound, newrate):
    sig, sr = sound
    if (sr == newrate):
        return sound
    numchannels = sig.shape[0]
    newsound = torchaudio.transforms.Resample(sr,newrate)(sig[:1,:])
    if (numchannels > 1):
        newsound2 = torchaudio.transforms.Resample(sr,newrate)(sig[1:,:])
        reig = torch.cat([newsound, newsound2])
    return(newsound)
    
# Function returns sound if in desired format (mono or stero) 
# Function returns mono if stero and stereo if mono by selecting only the first channel or duplicating it
# Outputs a tensor
def rechannel(sound, numchannels):
    sig, sr = torchaudio.load(sound)
    if (sig.shape[0] == numchannels):
        return sound
    if(numchannels == 1):
        newsound = sig[:1, :]
    else:
        newsond = torch.cat([sig,sig])
    return(newsound)
'''