dianskip/ai-voice-cloning
1
0
Fork 0
forked from camenduru/ai-voice-cloning
Code Pull Requests Activity

multichannel audio now report correct duration (surprised it took this long for me to source multichannel audio)

Browse Source
This commit is contained in:
mrq 2023-03-13 21:24:51 +00:00
parent 32d968a8cd
commit dadb1fca6b
1 changed files with 5 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
src/utils.py
View File

@ -519,7 +519,7 @@ def update_baseline_for_latents_chunks( voice ):
continue continue
metadata = torchaudio.info(f'{path}/{file}') metadata = torchaudio.info(f'{path}/{file}')
duration = metadata.num_channels * metadata.num_frames / metadata.sample_rate duration = metadata.num_frames / metadata.sample_rate
total_duration += duration total_duration += duration
total = total + 1 total = total + 1
@ -1079,7 +1079,7 @@ def validate_waveform( waveform, sample_rate, min_only=False ):
return "Waveform is empty" return "Waveform is empty"
num_channels, num_frames = waveform.shape num_channels, num_frames = waveform.shape
duration = num_channels * num_frames / sample_rate duration = num_frames / sample_rate
if duration < MIN_TRAINING_DURATION: if duration < MIN_TRAINING_DURATION:
return "Duration too short ({:.3f}s < {:.3f}s)".format(duration, MIN_TRAINING_DURATION) return "Duration too short ({:.3f}s < {:.3f}s)".format(duration, MIN_TRAINING_DURATION)
@ -1176,7 +1176,7 @@ def slice_dataset( voice, trim_silence=True, start_offset=0, end_offset=0, resul
result = results[filename] result = results[filename]
waveform, sample_rate = torchaudio.load(path) waveform, sample_rate = torchaudio.load(path)
num_channels, num_frames = waveform.shape num_channels, num_frames = waveform.shape
duration = num_channels * num_frames / sample_rate duration = num_frames / sample_rate
for segment in result['segments']: for segment in result['segments']:
file = filename.replace(".wav", f"_{pad(segment['id'], 4)}.wav") file = filename.replace(".wav", f"_{pad(segment['id'], 4)}.wav")
@ -1234,7 +1234,7 @@ def prepare_dataset( voice, use_segments, text_length, audio_length, normalize=F
continue continue
metadata = torchaudio.info(path) metadata = torchaudio.info(path)
duration = metadata.num_channels * metadata.num_frames / metadata.sample_rate duration = metadata.num_frames / metadata.sample_rate
if duration >= MAX_TRAINING_DURATION: if duration >= MAX_TRAINING_DURATION:
message = f"Audio too large, using segments: {filename}" message = f"Audio too large, using segments: {filename}"
print(message) print(message)
@ -1285,7 +1285,7 @@ def prepare_dataset( voice, use_segments, text_length, audio_length, normalize=F
culled = len(text) < text_length culled = len(text) < text_length
if not culled and audio_length > 0: if not culled and audio_length > 0:
num_channels, num_frames = waveform.shape num_channels, num_frames = waveform.shape
duration = num_channels * num_frames / sample_rate duration = num_frames / sample_rate
culled = duration < audio_length culled = duration < audio_length
# for when i add in a little treat ;), as it requires normalized text # for when i add in a little treat ;), as it requires normalized text