main.py

import os, sys, re, json
import argparse
#import whisper_timestamped as wt
from pdb import set_trace as b
from pprint import pprint as pp
from profanity_check import predict, predict_prob
from pydub import AudioSegment
from pydub.playback import play
from subprocess import Popen, PIPE


def parse_args(): 
    """ 
    """ 
    parser = argparse.ArgumentParser(
        description=('Tool to mute profanities in a song (source separation -> speech recognition -> profanity detection -> mask profanities -> re-mix)'),
        usage=('see <py main.py --help> or run as local web app with streamlit: <streamlit run main.py>')
    )

    parser.add_argument( 
        '-i', 
        '--input', 
        default=None, 
        nargs='?', 
        #required=True,
        help=("path to a mp3")
    )
    parser.add_argument( 
        '-m', 
        '--model', 
        default='small', 
        nargs='?', 
        help=("model used by whisper for speech recognition: tiny, small (default), medium or large")
    )
    parser.add_argument( 
        '-p', 
        '--play', 
        default=False, 
        action='store_true',
        help=("play output audio at the end")
    )
    parser.add_argument( 
        '-v', 
        '--verbose', 
        default=False, 
        action='store_true',
        help=("print transcribed text and detected profanities to screen")
    )
    return parser.parse_args()



def main(args, input_file=None, model_size=None, verbose=False, play_output=False):
    """
    """
    if not input_file:
        input_file = args.input

    if not model_size:
        model_size = args.model 

    if not verbose:
        verbose = args.verbose
    
    if not play_output:
        play_output = args.play

    # exit if input file not found
    if not os.path.isfile(input_file):
        print('Error: --input file not found')
        sys.exit()
    
    print(f'\nProcessing input file: {input_file}')

    # split audio into vocals + accompaniment
    print('Running source separation')
    stems_dir = source_separation(input_file)
    vocal_stem = os.path.join(stems_dir, 'vocals.wav')
    instr_stem = os.path.join(stems_dir, 'no_vocals.wav')
    print(f'Vocal stem written to: {vocal_stem}')

    # speech rec (audio->text)
    print('Transcribe vocal stem into text with word-level timestamps')
    #cmd = f'whisper_timestamped --task transcribe --model {model_size} {vocal_stem}'
    #stdout, stderr = Popen(cmd, stdout=PIPE, stderr=PIPE, shell=True, executable='/bin/bash').communicate()
    #text = json.loads('\n'.join(stdout.decode('utf8').split('\n')[1:]))

    import whisper_timestamped as wt
    audio = wt.load_audio(vocal_stem)
    model = wt.load_model(model_size, device='cpu')
    text = wt.transcribe(model, audio, language='en')
    
    if verbose:
        print('\nTranscribed text:')
        print(text['text']+'\n')
    
    # checking for profanities in text
    print('Run profanity detection on text')
    profanities = profanity_detection(text)
    if not profanities:
        print(f'No profanities found in {input_file} - exiting')
        sys.exit()
    if verbose:
        print('profanities found in text:')
        pp(profanities)
 
    # masking
    print('Mask profanities in vocal stem')
    vocals = mask_profanities(vocal_stem, profanities)

    # re-mixing
    print('Merge instrumentals stem and masked vocals stem')
    mix = AudioSegment.from_wav(instr_stem).overlay(vocals)

    # write mix to file
    outpath = input_file.replace('.mp3', '_masked.mp3').replace('.wav', '_masked.wav')
    if input_file.endswith('.wav'):
        mix.export(outpath, format="wav")
    elif input_file.endswith('.mp3'):
        mix.export(outpath, format="mp3")
    print(f'Mixed file written to: {outpath}')

    # play output
    if play_output:
        print('\nPlaying output...')
        play(mix)

    return outpath


def source_separation(inpath): 
    """
    Execute shell command to run demucs and pipe stdout/stderr back to python
    """
    cmd = f'demucs --two-stems=vocals --jobs 8 "{inpath}"'
    stdout, stderr = Popen(cmd, stdout=PIPE, stderr=PIPE, shell=True, executable='/bin/bash').communicate()
    stdout = stdout.decode('utf8')
    
    # exit if demucs error'd out
    if stderr:
        stderr = stderr.decode('utf-8').lower()
        if 'error' in stderr or 'not exist' in stderr:
            print(stderr.decode('utf8').split('\n')[0])
            sys.exit()

    # parse stems directory path from stdout and return it if successful
    stems_dir = ''.join(re.findall('/.*', stdout)).replace('.mp3','').replace('.wav','').replace('samples/','')
    if not os.path.isdir(stems_dir):
        print(f'Error: output stem directory "{stems_dir}" not found')
        sys.exit()

    return stems_dir


def profanity_detection(text):
    """
    """
    # detect profanities in text
    profs = []
    for segment in text['segments']:
        for word in segment['words']:
            #if word['confidence']<.25:
            #    print(word)
            text = word['text'].replace('.','').replace(',','').lower()

            # skip false positives
            if text in ['cancer', 'hell', 'junk', 'die', 'lame', 'freak', 'freaky', 'white', 'stink', 'shut', 'spit', 'mouth','orders','eat','clouds']:
                continue

            # assume anything returned by whisper with more than 1 * is profanity e.g n***a
            if '**' in text:
                profs.append(word)
                continue

            # add true negatives
            if text in ['bitchy', 'puss']:
                profs.append(word)
                continue

            # run profanity detection - returns 1 (True) or 0 (False)
            if predict([word['text']])[0]:
                profs.append(word)

    return profs


def mask_profanities(vocal_stem, profanities):
    """
    """
    # load vocal stem and mask profanities
    vocals = AudioSegment.from_wav(vocal_stem)
    for prof in profanities:
        mask   = vocals[prof['start']*1000:prof['end']*1000] # pydub works in milliseconds
        mask  -= 50 # reduce lvl by some dB (enough to ~mute it)
        #mask   = mask.silent(len(mask))
        #mask   = mask.fade_in(100).fade_out(100) # it prepends/appends fades so end up with longer mask
        start  = vocals[:prof['start']*1000]
        end    = vocals[prof['end']*1000:]
        #print(f"masking {prof['text']} from {prof['start']} to {prof['end']}")
        vocals = start + mask + end

    return vocals



if __name__ == "__main__":
    args = parse_args()
    
    if len(sys.argv)>1:
        main(args)
    else:
        import streamlit as st
        st.title('Saylss')
        model = st.selectbox('Choose model size:', ('tiny','small','medium','large'), index=1)
        uploaded_file = st.file_uploader("Choose input track", type=[".mp3",".wav"], accept_multiple_files=False)

        if uploaded_file is not None:
            
            # display input audio
            #st.text('Play input track:')
            audio_bytes_input = uploaded_file.read()
            st.audio(audio_bytes_input, format='audio/wav')
            
            # run code
            with st.spinner('Processing input audio...'):
                outpath = main(args, input_file=os.path.join('audio/samples',uploaded_file.name), model_size=model)
            
            # display output audio
            #st.text('Play output Track:')
            st.text('\nOutput:')
            audio_file = open(outpath, 'rb')
            audio_bytes = audio_file.read()
            st.audio(audio_bytes, format='audio/wav')