storytold/storyteller-ml
2
0
Fork 0
mirror of https://github.com/storytold/storyteller-ml.git synced 2026-10-15 00:03:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

Got subchunking for audio lower than 15 seconds to be chunked a bit further then cross faded.

Browse Source
This commit is contained in:
ArEnSc
2023-03-21 21:10:55 -04:00
parent 8bac589838
commit af65e49da6
1 changed files with 81 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
voice_conversion/so-vits-svc/inference_main.py
+81 -14
View File
@@ -16,6 +16,22 @@ logging.getLogger('numba').setLevel(logging.WARNING)
chunks_dict = infer_tool.read_temp("inference/chunks_temp.json")
def stitch_wav_tensors_with_crossfade(wav_tensor1,wav_tensor2,sample_rate:int):
# Load two audio signals as NumPy tensors
# Compute the length of the cross-fade window in samples
fade_len = int(0.01 * sample_rate)
# Create a cross-fade window
fade_window = np.hanning(2*fade_len)
# Concatenate the signals with cross-fade
audio_out = np.concatenate((wav_tensor1[:-fade_len],
wav_tensor1[-fade_len:] * fade_window[:fade_len] + wav_tensor2[:fade_len] * fade_window[fade_len:],
wav_tensor2[fade_len:]))
#sf.write("audio_out.wav", audio_out, sample_rate)
return audio_out
def main():
import argparse
@@ -75,21 +91,72 @@ def main():
if slice_tag:
print('jump empty segment')
_audio = np.zeros(length)
else:
# padd
elif length >= 15:
# pad
pad_len = int(audio_sr * pad_seconds)
data = np.concatenate([np.zeros([pad_len]), data, np.zeros([pad_len])])
raw_path = io.BytesIO()
soundfile.write(raw_path, data, audio_sr, format="wav")
raw_path.seek(0)
out_audio, out_sr = svc_model.infer(spk, tran, raw_path,
cluster_infer_ratio=cluster_infer_ratio,
auto_predict_f0=auto_predict_f0,
noice_scale=noice_scale
)
_audio = out_audio.cpu().numpy()
pad_len = int(svc_model.target_sample * pad_seconds)
_audio = _audio[pad_len:-pad_len]
# if chunk size is too large more than 15 seconds break it into smaller pieces because sliced by large amounts
# final piece
final_cross_faded_chunk_wav_tensor = None
# chunk into smaller pieces cross fade
# Set the length of each chunk in seconds
chunk_length = 15
# Calculate the number of samples in each chunk
chunk_samples = chunk_length * audio_sr
# Calculate the total number of chunks
total_chunks = int(len(data) / chunk_samples)
# Calculate the number of remaining samples
remaining_samples = len(data) - (total_chunks * chunk_samples)
# Add an additional chunk for any remaining samples
if remaining_samples > 0:
total_chunks += 1
# Slice the WAV tensor into chunks
chunks = [data[i * chunk_samples:min((i + 1) * chunk_samples, len(data))] for i in range(total_chunks)]
for chunk in chunks:
# prepare for inference
data = np.concatenate([np.zeros([pad_len]), chunk, np.zeros([pad_len])])
# Testing writes to the audio to do a match up.
soundfile.write(raw_path, data, audio_sr, format="wav")
raw_path.seek(0)
out_audio, out_sr = svc_model.infer(spk, tran, raw_path,
cluster_infer_ratio=cluster_infer_ratio,
auto_predict_f0=auto_predict_f0,
noice_scale=noice_scale
)
#trim padded audio by new sample length and cross fade
pad_len = int(svc_model.target_sample * pad_seconds)
_audio = _audio[pad_len:-pad_len]
_audio = out_audio.cpu().numpy()
if final_cross_faded_chunk_wav_tensor == None:
final_cross_faded_chunk_wav_tensor = _audio
else:
final_cross_faded_chunk_wav_tensor = stitch_wav_tensors_with_crossfade(wav_tensor1=final_cross_faded_chunk_wav_tensor,wav_tensor2=_audio,sample_rate=svc_model.target_sample)
_audio = final_cross_faded_chunk_wav_tensor
else:
data = np.concatenate([np.zeros([pad_len]), data, np.zeros([pad_len])])
raw_path = io.BytesIO()
soundfile.write(raw_path, data, audio_sr, format="wav")
raw_path.seek(0)
out_audio, out_sr = svc_model.infer(spk, tran, raw_path,
cluster_infer_ratio=cluster_infer_ratio,
auto_predict_f0=auto_predict_f0,
noice_scale=noice_scale
)
_audio = out_audio.cpu().numpy()
#trim padded audio by new sample length
pad_len = int(svc_model.target_sample * pad_seconds)
# trim padded both sides with new sample ra
_audio = _audio[pad_len:-pad_len]
audio.extend(list(infer_tool.pad_array(_audio, length)))
key = "auto" if auto_predict_f0 else f"{tran}key"