skip_music.py

from pydub import AudioSegment
from google.cloud import speech_v1p1beta1 as speech
from google.cloud.speech_v1p1beta1 import enums

AUDIO_FILE_PATH_LOCAL = '<PATH TO YOUR FILE WHERE YOU WANT TO REMOVE MUSIC>'
AUDIO_FILE_PATH_GCS = '<PATH TO THE SAME FILE IN GCS IF YOU USE THE FIRST SOLUTION>'
OUTPUT_FILE_PATH_LOCAL = '<PATH TO SAVE THE OUTPUT FILE>'

# Prepare audio files
show_without_music = AudioSegment.empty()
audio_segment = AudioSegment.from_mp3(AUDIO_FILE_PATH_LOCAL)

client = speech.SpeechClient()
speech_config = speech.types.RecognitionConfig(
    encoding=enums.RecognitionConfig.AudioEncoding.MP3,
    sample_rate_hertz=44100,
    language_code="fr-FR",
    audio_channel_count=2,
    enable_word_time_offsets=True,
)

#-------First solution: send the whole file to STT (needs a copy in GCS)-----------

# Run STT

recognition_audio = {'uri': AUDIO_FILE_PATH_GCS}
operation = client.long_running_recognize(speech_config, recognition_audio)
print(u"Waiting Speech To Text to complete...")
response = operation.result(timeout=2000)

# Interpret results

first_text_start = None
last_text_end = None

for result in response.results:
    print(result.alternatives[0])
    best_alternative = result.alternatives[0]
    if best_alternative.confidence > 0.9:
        nb_words = len(best_alternative.words)
        if first_text_start is None: # explicitely compares to None because it can be 0 which is considered false
            first_text_start = best_alternative.words[0].start_time.seconds
        last_text_end = result.alternatives[0].words[nb_words-1].end_time.seconds
    else:
        if first_text_start is not None:
            show_without_music += audio_segment[first_text_start*1000:last_text_end*1000 + 1000]
            first_text_start = None
            last_text_end = None
        else:
            continue

if first_text_start:
    show_without_music += audio_segment[first_text_start * 1000:last_text_end * 1000 + 1000]

# Write output file to disk
show_without_music.export(OUTPUT_FILE_PATH_LOCAL, format="mp3")

# Use this to generate smaller files
# audio_segment = AudioSegment.from_mp3(AUDIO_FILE_PATH_LOCAL)
# audio_segment = audio_segment[0:930*1000]
# audio_segment.export(OUTPUT_FILE_PATH_LOCAL, format="mp3")

#-------Second solution: cut file into 5 seconds chunks and send to the api-----

# CHUNKS_SECONDS = 5
#
# speech_config.enable_word_time_offsets = False
# output = io.BytesIO()
# nb_splits = int(len(audio_segment)/1000/CHUNKS_SECONDS)+1
# for i in range(nb_splits):
#     start_second, end_second = i*CHUNKS_SECONDS, (i+1)*CHUNKS_SECONDS
#     print("processing second %d to second %d" % (start_second, end_second))
#     fivesec_audio = audio_segment[start_second*1000:end_second*1000]
#     fivesec_audio.export(output, format="mp3")
#     recognition_audio = {'content': output.read()}
#     output.flush()
#     response = client.recognize(speech_config, recognition_audio)
#     print(response.results)
#     if len(response.results) > 0:
#         print("Speech")
#         show_without_music += fivesec_audio
#     else:
#         print("Music")
#
#
# show_without_music.export(OUTPUT_FILE_PATH_LOCAL, format="mp3")
# print("File without music written at %s" % OUTPUT_FILE_PATH_LOCAL)