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

WIP: pitch extractor

Browse Source
This commit is contained in:
w-okada 2023-07-15 05:30:07 +09:00
parent 6a09338af5
commit 02eadeeaa3
2 changed files with 14 additions and 67 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
server/voice_changer/DiffusionSVC/pipeline/Pipeline.py
View File

@ -17,8 +17,6 @@ from voice_changer.RVC.embedder.Embedder import Embedder
from voice_changer.common.VolumeExtractor import VolumeExtractor
from torchaudio.transforms import Resample
from voice_changer.utils.Timer import Timer
class Pipeline(object):
embedder: Embedder
@ -144,23 +142,15 @@ class Pipeline(object):
silence_front=silence_front,
)
pitch = torch.tensor(pitch[-n_frames:], device=self.device).unsqueeze(0).long()
pitch = torch.tensor(pitch[-n_frames:], device=self.device).unsqueeze(0).long()
except IndexError as e: # NOQA
# print(e)
raise NotEnoughDataExtimateF0()
print("[EMBEDDER EXTRACT:audio:4:]", audio_t.shape)
# f0 = self.f0ex.extract_f0(audio_pad, key=4, sr=44100)
# print("[Pitch_f0]", f0)
# tensor型調整
feats = audio16k.squeeze()
if feats.dim() == 2: # double channels
feats = feats.mean(-1)
feats = feats.view(1, -1)
print("[EMBEDDER EXTRACT:audio:5:]", audio_t.shape)
print("[EMBEDDER EXTRACT:::]", feats.shape)
# embedding
with autocast(enabled=self.isHalf):
@ -177,39 +167,6 @@ class Pipeline(object):
raise e
feats = F.interpolate(feats.permute(0, 2, 1), size=int(n_frames), mode='nearest').permute(0, 2, 1)
if protect < 0.5:
feats0 = feats.clone()
# # ピッチサイズ調整
# p_len = audio_pad.shape[0] // self.window
# feats_len = feats.shape[1]
# if feats.shape[1] < p_len:
# p_len = feats_len
# pitch = pitch[:, :feats_len]
# pitchf = pitchf[:, :feats_len]
# pitch = pitch[:, -feats_len:]
# pitchf = pitchf[:, -feats_len:]
# p_len = torch.tensor([feats_len], device=self.device).long()
# print("----------plen::1:", p_len)
# pitchの推定が上手くいかない(pitchf=0)場合、検索前の特徴を混ぜる
# pitchffの作り方の疑問はあるが、本家通りなので、このまま使うことにする。
# https://github.com/w-okada/voice-changer/pull/276#issuecomment-1571336929
# if protect < 0.5:
# pitchff = pitchf.clone()
# pitchff[pitchf > 0] = 1
# pitchff[pitchf < 1] = protect
# pitchff = pitchff.unsqueeze(-1)
# feats = feats * pitchff + feats0 * (1 - pitchff)
# feats = feats.to(feats0.dtype)
# # apply silent front for inference
# if type(self.inferencer) in [OnnxRVCInferencer, OnnxRVCInferencerNono]:
# npyOffset = math.floor(silence_front * 16000) // 360 # 160x2 = 360
# feats = feats[:, npyOffset * 2 :, :] # NOQA
# 推論実行
try:
with torch.no_grad():

36
server/voice_changer/DiffusionSVC/pitchExtractor/HarvestPitchExtractor.py
View File

@ -1,6 +1,5 @@
import pyworld
import numpy as np
import scipy.signal as signal
from const import PitchExtractorType
import torch
from voice_changer.RVC.pitchExtractor.PitchExtractor import PitchExtractor
@ -11,40 +10,31 @@ class HarvestPitchExtractor(PitchExtractor):
def __init__(self):
super().__init__()
self.pitchExtractorType: PitchExtractorType = "harvest"
self.f0_min = 50
self.f0_max = 1100
self.sapmle_rate = 16000
self.uv_interp = True
def extract(self, audio: torch.Tensor, pitchf, f0_up_key, sr, window, silence_front=0):
audio = audio.detach().cpu().numpy()
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)
f0 = self.extract2(audio, uv_interp=True, hop_size=window, silence_front=silence_front)
f0 = f0 * 2 ** (float(f0_up_key) / 12)
# pitchf[-f0.shape[0]:] = f0[:pitchf.shape[0]]
return f0
def extract2(self, audio, uv_interp, hop_size: int, silence_front=0): # audio: 1d numpy array
n_frames = int(len(audio) // hop_size) + 1
start_frame = int(silence_front * 16000 / hop_size)
real_silence_front = start_frame * hop_size / 16000
audio = audio[int(np.round(real_silence_front * 16000)):]
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, _ = pyworld.harvest(
audio.astype('double'),
16000,
f0_floor=50,
f0_ceil=1100,
frame_period=(1000 * hop_size / 16000))
f0 = np.pad(f0.astype('float'), (start_frame, n_frames - len(f0) - start_frame))
frame_period=(1000 * window / self.sapmle_rate))
pitchf[-f0.shape[0]:] = f0[:pitchf.shape[0]]
f0 = pitchf
if uv_interp:
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 < 50] = 50
f0 = f0 * 2 ** (float(f0_up_key) / 12)
return f0