def vad_segment_generator(wavFile, aggressiveness):
logging.debug("Caught the wav file @: %s" % (wavFile))
audio, sample_rate, audio_length = wavSplit.read_wave(wavFile)
assert sample_rate == 16000, "Only 16000Hz input WAV files are supported for now!"
#vad = webrtcvad.Vad(int(aggressiveness))
frames = list(wavSplit.frame_generator(20000, audio, sample_rate))
return frames
def vad_segment_generator(wavFile, aggressiveness):
logging.debug("Caught the wav file @: %s" % (wavFile))
audio, sample_rate, audio_length = wavSplit.read_wave(wavFile)
vad = webrtcvad.Vad(int(aggressiveness))
frames = wavSplit.frame_generator(10, audio, sample_rate)
frames = list(frames)
segments = wavSplit.vad_collector(sample_rate, 10, 300, vad, frames)
return segments, sample_rate, audio_length
def vad_segment_generator(wavFile, aggressiveness, frame_duration_ms=30, padding_duration_ms=300):
logging.debug("Caught the wav file @: %s" % (wavFile))
audio, sample_rate, audio_length = wavSplit.read_wave(wavFile)
assert sample_rate == 16000, "Only 16000Hz input WAV files are supported for now!"
vad = webrtcvad.Vad(int(aggressiveness))
frames = wavSplit.frame_generator(frame_duration_ms, audio, sample_rate)
frames = list(frames)
segments = wavSplit.vad_collector(sample_rate, frame_duration_ms, padding_duration_ms, vad, frames)
return [segment for segment in segments], sample_rate, audio_length
def vad_segment_generator(wavFile, aggressiveness):
logging.debug("Caught the wav file @: %s" % (wavFile))
audio, sample_rate, audio_length = wavSplit.read_wave(wavFile)
assert sample_rate == 16000, "Only 16000Hz input WAV files are supported for now!"
vad = webrtcvad.Vad(int(aggressiveness))
frames = wavSplit.frame_generator(30, audio, sample_rate)
frames = list(frames)
segments = wavSplit.vad_collector(sample_rate, 30, 300, vad, frames)
return segments, sample_rate, audio_length
def vad_segment_generator(wavFile, aggressiveness, model_sample_rate):
logging.debug("Caught the wav file @: %s" % (wavFile))
audio, sample_rate, audio_length = wavSplit.read_wave(wavFile)
assert sample_rate == model_sample_rate, \
"Audio sample rate must match sample rate of used model: {}Hz".format(model_sample_rate)
vad = webrtcvad.Vad(int(aggressiveness))
frames = wavSplit.frame_generator(30, audio, sample_rate)
frames = list(frames)
segments = wavSplit.vad_collector(sample_rate, 30, 300, vad, frames)
return segments, sample_rate, audio_length
def vad_segment_generator(wav_file, aggressiveness):
"""
Generate VAD segments. Filters out non-voiced audio frames.
:param wav_file: Input wav file to run VAD on.0
:param aggressiveness: How aggressive filtering out non-speech is (between 0 and 3)
:return: Returns tuple of
segments: a bytearray of multiple smaller audio frames
(The longer audio split into multiple smaller one's)
sample_rate: Sample rate of the input audio file
audio_length: Duration of the input audio file
"""
logging.debug("Caught the wav file @: %s" % wav_file)
audio, sample_rate, audio_length = wavSplit.read_wave(wav_file)
assert sample_rate == 16000, "Only 16000Hz input WAV files are supported for now!"
vad = webrtcvad.Vad(int(aggressiveness))
frames = wavSplit.frame_generator(30, audio, sample_rate)
frames = list(frames)
segments = wavSplit.vad_collector(sample_rate, 30, 300, 0.5, vad, frames)
return segments, sample_rate, audio_length
ds.enableExternalScorer(scorer)
# ds === deep speech model
# load and pre-process audio
audio_file = os.path.join(os.getcwd(),
"data/testing/audio/2830-3980-0043.wav")
# audio_file = os.path.join(os.getcwd(),"data/testing/audio/my_name_is_jamie.wav")
# audio_file = os.path.join(os.getcwd(),"data/testing/audio/hello_liv.wav")
aggressiveness = 0
print("Reading and processing: {}".format(audio_file))
audio, sample_rate, audio_length = wavSplit.read_wave(audio_file)
assert sample_rate == 16000, "Only 16000Hz input WAV files are supported for now!"
vad = webrtcvad.Vad(int(aggressiveness))
frames = wavSplit.frame_generator(30, audio, sample_rate)
frames = list(frames)
segments = wavSplit.vad_collector(sample_rate, 30, 300, vad, frames)
# we now have the data in the following segments
# segments, sample_rate, audio_length
print("we have {} frames".format(len(frames)))
start = time.time()
for i, segment in enumerate(segments):
# Run deepspeech on the chunk that just completed VAD
print("Processing chunk %002d" % (i, ))
audio = np.frombuffer(segment, dtype=np.int16)
# Run Deepspeech
print('Running inference...')
output = ds.stt(audio)
print("Transcript: %s" % output)
def transcribe_file(audio_path, ds, threadcount):
audio, sample_rate, audio_length = wavSplit.read_wave(audio_path)
segments = list(wavSplit.frame_generator(CHUNK_SIZE, audio, sample_rate))
print("Beginning to process generated file chunks")
inference_time = time.time()
# make a set of queues for upstream and downstream communication
WriteQueue = queue.Queue()
ReadQueue = queue.Queue()
workers = []
threadcount = int(threadcount)
for i in range(threadcount):
x = ChunkWorker(
WriteQueue, ReadQueue,
ds) # all chunks get the same queues and inference model
x.start()
workers.append(x)
if len(segments) <= 1:
print("SINGLE SEGMENT LENGTH IDENTIFIED", file=sys.stderr, flush=True)
# this can occur when the chunk size is larger than the audio file, resulting in no chunking
# we need to do the transcription manually
if len(segments) == 0:
audio_file = wave.open(audio_path, 'rb')
output = ds.sttWithMetadata(audio_file.bytes, 1) # Run Deepspeech
return json.dumps(output)
print("Workers started...", file=sys.stderr, flush=True)
for i, segment in enumerate(segments):
print("Writing segment num {}".format(i), file=sys.stderr, flush=True)
WriteQueue.put({'chunk': segment, 'index': i})
print("All Chunks sent...", file=sys.stderr, flush=True)
for i in range(threadcount):
print("stopping worker {}".format(i), file=sys.stderr, flush=True)
WriteQueue.put({"index": -1}) # send a sentinel value to all threads
for i in range(threadcount):
workers[i].join() # wait for all threads to join
print(" worker {} has joined".format(i), file=sys.stderr, flush=True)
processed_chunks = []
for ele in list(ReadQueue.queue):
print(ele, file=sys.stderr, flush=True)
processed_chunks.append(ele)
# each thread will send both the inference result
# as well as the chunk id which is used to sort
# the ReadQueue, allowing for asynchronous processing
processed_chunks.sort(key=lambda p: p.get('index'))
# because each chunk is processed discretely, each word will have a time
# value within the range of zero and CHUNK_SIZE in seconds represented by a float value.
# To accurately associate each word with its proper time within the media, and not within
# the range described above, we need to add an offset to all tokens in each segment
# - excluding the zeroth segment as it needs no adjustment.
currentTime = 0.0
i = 0
adjusted_tokens = []
offset = CHUNK_SIZE / 1000
current_offset = 0
if len(processed_chunks) > 1:
while i < len(processed_chunks):
current_item = processed_chunks[i].get('result')
if len(
current_item
) != 0: # if we get a chunk that has no speech within it, such as instrumental music, ignore it
j = 0
while j < len(current_item):
current_word = current_item[j]
current_word[
'time'] = current_word['time'] + current_offset
j = j + 1
adjusted_tokens.append(current_word)
i = i + 1
current_offset = current_offset + offset
else:
print("done with file; took{}".format(time.time() - inference_time),
file=sys.stderr,
flush=True)
return json.dumps(processed_chunks[0].get('result'))
print("done with file; took{}".format(time.time() - inference_time),
file=sys.stderr,
flush=True)
return json.dumps(adjusted_tokens)
