1berry/wokada-voice-changer
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

WIP: cprep onnx

Browse Source
This commit is contained in:
w-okada 2023-07-15 06:07:33 +09:00
parent 485a747d55
commit 333e66786d
3 changed files with 29 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

48
server/voice_changer/DiffusionSVC/pitchExtractor/CrepeOnnxPitchExtractor.py
View File

@ -4,6 +4,7 @@ from voice_changer.RVC.deviceManager.DeviceManager import DeviceManager
from voice_changer.RVC.pitchExtractor.PitchExtractor import PitchExtractor
import onnxruntime
from voice_changer.RVC.pitchExtractor import onnxcrepe
import torch
class CrepeOnnxPitchExtractor(PitchExtractor):
@ -19,30 +20,26 @@ class CrepeOnnxPitchExtractor(PitchExtractor):
self.onnx_session = onnxruntime.InferenceSession(
file, providers=onnxProviders, provider_options=onnxProviderOptions
)
self.f0_min = 50
self.f0_max = 1100
self.sapmle_rate = 16000
self.uv_interp = True
def extract(self, audio, pitchf, f0_up_key, sr, window, silence_front=0):
n_frames = int(len(audio) // window) + 1
start_frame = int(silence_front * sr / window)
real_silence_front = start_frame * window / sr
def extract(self, audio: torch.Tensor, pitch, f0_up_key, window, silence_front=0):
start_frame = int(silence_front * self.sapmle_rate / window)
real_silence_front = start_frame * window / self.sapmle_rate
audio = audio[int(np.round(real_silence_front * self.sapmle_rate)):]
silence_front_offset = int(np.round(real_silence_front * sr))
audio = audio[silence_front_offset:]
f0_min = 50
f0_max = 1100
f0_mel_min = 1127 * np.log(1 + f0_min / 700)
f0_mel_max = 1127 * np.log(1 + f0_max / 700)
precision = 10.0
precision = (1000 * window / self.sapmle_rate)
audio_num = audio.cpu()
onnx_f0, onnx_pd = onnxcrepe.predict(
self.onnx_session,
audio_num,
sr,
self.sapmle_rate,
precision=precision,
fmin=f0_min,
fmax=f0_max,
fmin=self.f0_min,
fmax=self.f0_max,
batch_size=256,
return_periodicity=True,
decoder=onnxcrepe.decode.weighted_argmax,
@ -53,14 +50,15 @@ class CrepeOnnxPitchExtractor(PitchExtractor):
f0[pd < 0.1] = 0
f0 = f0.squeeze()
pitch[-f0.shape[0]:] = f0[:pitch.shape[0]]
f0 = pitch
f0 *= pow(2, f0_up_key / 12)
pitchf[-f0.shape[0]:] = f0[:pitchf.shape[0]]
f0bak = pitchf.copy()
f0_mel = 1127.0 * np.log(1.0 + f0bak / 700.0)
f0_mel = np.clip(
(f0_mel - f0_mel_min) * 254.0 / (f0_mel_max - f0_mel_min) + 1.0, 1.0, 255.0
)
pitch_coarse = f0_mel.astype(int)
if self.uv_interp:
uv = f0 == 0
if len(f0[~uv]) > 0:
f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
f0[f0 < self.f0_min] = self.f0_min
return pitch_coarse, pitchf
f0 = f0 * 2 ** (float(f0_up_key) / 12)
return f0

2
server/voice_changer/DiffusionSVC/pitchExtractor/CrepePitchExtractor.py
View File

@ -24,7 +24,7 @@ class CrepePitchExtractor(PitchExtractor):
start_frame = int(silence_front * self.sapmle_rate / window)
real_silence_front = start_frame * window / self.sapmle_rate
audio = audio[int(np.round(real_silence_front * self.sapmle_rate)):]
f0, pd = torchcrepe.predict(
audio.unsqueeze(0),
self.sapmle_rate,

9
server/voice_changer/DiffusionSVC/pitchExtractor/PitchExtractorManager.py
View File

@ -1,5 +1,6 @@
from typing import Protocol
from const import PitchExtractorType
from voice_changer.DiffusionSVC.pitchExtractor.CrepeOnnxPitchExtractor import CrepeOnnxPitchExtractor
from voice_changer.DiffusionSVC.pitchExtractor.CrepePitchExtractor import CrepePitchExtractor
from voice_changer.DiffusionSVC.pitchExtractor.DioPitchExtractor import DioPitchExtractor
from voice_changer.DiffusionSVC.pitchExtractor.HarvestPitchExtractor import HarvestPitchExtractor
@ -32,10 +33,10 @@ class PitchExtractorManager(Protocol):
return DioPitchExtractor()
elif pitchExtractorType == "crepe":
return CrepePitchExtractor()
# elif pitchExtractorType == "crepe_tiny":
# return CrepeOnnxPitchExtractor(pitchExtractorType, cls.params.crepe_onnx_tiny, gpu)
# elif pitchExtractorType == "crepe_full":
# return CrepeOnnxPitchExtractor(pitchExtractorType, cls.params.crepe_onnx_full, gpu)
elif pitchExtractorType == "crepe_tiny":
return CrepeOnnxPitchExtractor(pitchExtractorType, cls.params.crepe_onnx_tiny, gpu)
elif pitchExtractorType == "crepe_full":
return CrepeOnnxPitchExtractor(pitchExtractorType, cls.params.crepe_onnx_full, gpu)
else:
# return hubert as default
raise RuntimeError(