when creating the train/validatio datasets, use segments if the main audio's duration is too long, and slice to make the segments if they don't exist

This commit is contained in:
mrq 2023-03-13 04:26:00 +00:00
parent 0cf9db5e69
commit ee1b048d07

View File

@ -51,6 +51,9 @@ LEARNING_RATE_SCHEDULE = [ 9, 18, 25, 33 ]
RESAMPLERS = {}
MIN_TRAINING_DURATION = 0.6
MAX_TRAINING_DURATION = 11.6097505669
args = None
tts = None
tts_loading = False
@ -62,6 +65,9 @@ training_state = None
current_voice = None
def resample( waveform, input_rate, output_rate=44100 ):
# mono-ize
waveform = torch.mean(waveform, dim=0, keepdim=True)
if input_rate == output_rate:
return waveform, output_rate
@ -1066,18 +1072,19 @@ def whisper_transcribe( file, language=None ):
result['segments'].append(reparsed)
return result
def validate_waveform( waveform, sample_rate ):
def validate_waveform( waveform, sample_rate, min_only=False ):
if not torch.any(waveform < 0):
return "Waveform is empty"
num_channels, num_frames = waveform.shape
duration = num_channels * num_frames / sample_rate
if duration < 0.6:
return "Duration too short ({:.3f} < 0.6s)".format(duration)
if duration < MIN_TRAINING_DURATION:
return "Duration too short ({:.3f}s < {:.3f}s)".format(duration, MIN_TRAINING_DURATION)
if duration > 11:
return "Duration too long (11s < {:.3f})".format(duration)
if not min_only:
if duration > MAX_TRAINING_DURATION:
return "Duration too long ({:.3f}s < {:.3f}s)".format(MAX_TRAINING_DURATION, duration)
return
@ -1122,7 +1129,24 @@ def transcribe_dataset( voice, language=None, skip_existings=False, progress=Non
return f"Processed dataset to: {indir}"
def slice_dataset( voice, trim_silence=True, start_offset=0, end_offset=0 ):
def slice_waveform( waveform, sample_rate, start, end, trim ):
start = int(start * sample_rate)
end = int(end * sample_rate)
if start < 0:
start = 0
if end >= waveform.shape[-1]:
end = waveform.shape[-1] - 1
sliced = waveform[:, start:end]
error = validate_waveform( sliced, sample_rate, min_only=True )
if trim and not error:
sliced = torchaudio.functional.vad( sliced, sample_rate )
return sliced, error
def slice_dataset( voice, trim_silence=True, start_offset=0, end_offset=0, results=None ):
indir = f'./training/{voice}/'
infile = f'{indir}/whisper.json'
messages = []
@ -1130,6 +1154,7 @@ def slice_dataset( voice, trim_silence=True, start_offset=0, end_offset=0 ):
if not os.path.exists(infile):
raise Exception(f"Missing dataset: {infile}")
if results is None:
results = json.load(open(infile, 'r', encoding="utf-8"))
files = 0
@ -1140,37 +1165,35 @@ def slice_dataset( voice, trim_silence=True, start_offset=0, end_offset=0 ):
path = f'./training/{voice}/{filename}'
if not os.path.exists(path):
messages.append(f"Missing source audio: {filename}")
message = f"Missing source audio: {filename}"
print(message)
messages.append(message)
continue
files += 1
result = results[filename]
waveform, sample_rate = torchaudio.load(path)
num_channels, num_frames = waveform.shape
duration = num_channels * num_frames / sample_rate
for segment in result['segments']:
start = int((segment['start'] + start_offset) * sample_rate)
end = int((segment['end'] + end_offset) * sample_rate)
if start < 0:
start = 0
if end >= waveform.shape[-1]:
end = waveform.shape[-1] - 1
sliced = waveform[:, start:end]
file = filename.replace(".wav", f"_{pad(segment['id'], 4)}.wav")
if trim_silence:
sliced = torchaudio.functional.vad( sliced, sample_rate )
sliced, sample_rate = resample( sliced, sample_rate, 22050 )
torchaudio.save(f"{indir}/audio/{file}", sliced, sample_rate)
sliced, error = slice_waveform( waveform, sample_rate, segment['start'] + start_offset, segment['end'] + end_offset, trim_silence )
if error:
message = f"{error}, skipping... {file}"
print(message)
messages.append(message)
continue
sliced, _ = resample( sliced, sample_rate, 22050 )
torchaudio.save(f"{indir}/audio/{file}", sliced, 22050)
segments +=1
messages.append(f"Sliced segments: {files} => {segments}.")
return "\n".join(messages)
def prepare_dataset( voice, use_segments, text_length, audio_length ):
def prepare_dataset( voice, use_segments, text_length, audio_length, normalize=True ):
indir = f'./training/{voice}/'
infile = f'{indir}/whisper.json'
messages = []
@ -1187,31 +1210,67 @@ def prepare_dataset( voice, use_segments, text_length, audio_length ):
for filename in results:
result = results[filename]
segments = result['segments'] if use_segments else [{'text': result['text']}]
for segment in segments:
file = filename.replace(".wav", f"_{pad(segment['id'], 4)}.wav") if use_segments else filename
path = f'{indir}/audio/{file}'
use_segment = use_segments
# check if unsegmented audio exceeds 11.6s
if not use_segment:
path = f'{indir}/audio/{filename}'
if not os.path.exists(path):
messages.append(f"Missing source audio: {file}")
messages.append(f"Missing source audio: {filename}")
continue
metadata = torchaudio.info(path)
duration = metadata.num_channels * metadata.num_frames / metadata.sample_rate
if duration >= MAX_TRAINING_DURATION:
message = f"Audio too large, using segments: {filename}"
print(message)
messages.append(message)
use_segment = True
segments = result['segments'] if use_segment else [{'text': result['text']}]
for segment in segments:
file = filename.replace(".wav", f"_{pad(segment['id'], 4)}.wav") if use_segment else filename
path = f'{indir}/audio/{file}'
# segment when needed
if not os.path.exists(path):
tmp_results = {}
tmp_results[filename] = result
print(f"Audio not segmented, segmenting: {filename}")
message = slice_dataset( voice, results=tmp_results )
print(message)
messages = messages + message.split("\n")
if not os.path.exists(path):
message = f"Missing source audio: {file}"
print(message)
messages.append(message)
continue
text = segment['text'].strip()
normalized_text = text
if len(text) > 200:
messages.append(f"[{file}] Text length too long (200 < {len(text)}), skipping...")
message = f"Text length too long (200 < {len(text)}), skipping... {file}"
print(message)
messages.append(message)
waveform, sample_rate = torchaudio.load(path)
num_channels, num_frames = waveform.shape
duration = num_channels * num_frames / sample_rate
error = validate_waveform( waveform, sample_rate )
if error:
messages.append(f"[{file}]: {error}, skipping...")
message = f"{error}, skipping... {file}"
print(message)
messages.append(message)
continue
culled = len(text) < text_length
if not culled and audio_length > 0:
num_channels, num_frames = waveform.shape
duration = num_channels * num_frames / sample_rate
culled = duration < audio_length
# lines['training' if not culled else 'validation'].append(f'audio/{file}|{text}|{normalized_text}')
lines['training' if not culled else 'validation'].append(f'audio/{file}|{text}')
training_joined = "\n".join(lines['training'])