import torchcrepe import torch import numpy as np from const import EnumPitchExtractorTypes from voice_changer.RVC.pitchExtractor.PitchExtractor import PitchExtractor class CrepePitchExtractor(PitchExtractor): pitchExtractorType: EnumPitchExtractorTypes = EnumPitchExtractorTypes.crepe def __init__(self): super().__init__() if torch.cuda.is_available(): self.device = torch.device("cuda:" + str(torch.cuda.current_device())) else: self.device = torch.device("cpu") def extract(self, audio, 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 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) f0, pd = torchcrepe.predict( audio.unsqueeze(0), sr, hop_length=window, fmin=f0_min, fmax=f0_max, # model="tiny", model="full", batch_size=256, decoder=torchcrepe.decode.weighted_argmax, device=self.device, return_periodicity=True, ) f0 = torchcrepe.filter.median(f0, 3) # 本家だとmeanですが、harvestに合わせmedianフィルタ pd = torchcrepe.filter.median(pd, 3) f0[pd < 0.1] = 0 f0 = f0.squeeze() f0 = torch.nn.functional.pad( f0, (start_frame, n_frames - f0.shape[0] - start_frame) ) f0 *= pow(2, f0_up_key / 12) f0bak = f0.detach().cpu().numpy() f0_mel = 1127.0 * torch.log(1.0 + f0 / 700.0) f0_mel = torch.clip( (f0_mel - f0_mel_min) * 254.0 / (f0_mel_max - f0_mel_min) + 1.0, 1.0, 255.0 ) f0_coarse = f0_mel.round().detach().cpu().numpy().astype(int) return f0_coarse, f0bak