vc.py

import parselmouth
import numpy as np
import torch
from time import time as ttime
from torch.nn import functional as F


class VC(object):
    def __init__(self, tgt_sr, device, is_half, x_pad, x_query, x_center, x_max):
        self.sr = 16000  # hubert输入采样率
        self.window = 160  # 每帧点数
        self.t_pad = self.sr * x_pad  # 每条前后pad时间
        self.t_pad_tgt = tgt_sr * x_pad
        self.t_pad2 = self.t_pad * 2
        self.t_query = self.sr * x_query  # 查询切点前后查询时间
        self.t_center = self.sr * x_center  # 查询切点位置
        self.t_max = self.sr * x_max  # 免查询时长阈值
        self.device = device
        self.is_half = is_half
        self.x_pad = x_pad
        self.x_query = x_query
        self.x_center = x_center
        self.x_max = x_max

    def get_f0(self, x, p_len, f0_up_key, f0_method, inp_f0=None):
        time_step = self.window / self.sr * 1000
        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)
        if f0_method == "pm":
            f0 = (
                parselmouth.Sound(x, self.sr)
                .to_pitch_ac(
                    time_step=time_step / 1000,
                    voicing_threshold=0.6,
                    pitch_floor=f0_min,
                    pitch_ceiling=f0_max,
                )
                .selected_array["frequency"]
            )
            pad_size = (p_len - len(f0) + 1) // 2
            if pad_size > 0 or p_len - len(f0) - pad_size > 0:
                f0 = np.pad(
                    f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant"
                )
        f0 *= pow(2, f0_up_key / 12)
        # with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
        tf0 = self.sr // self.window  # 每秒f0点数
        if inp_f0 is not None:
            delta_t = np.round(
                (inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
            ).astype("int16")
            replace_f0 = np.interp(
                list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]
            )
            shape = f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)].shape[0]
            f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)] = replace_f0[
                :shape
            ]
        # with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
        f0bak = f0.copy()
        f0_mel = 1127 * np.log(1 + f0 / 700)
        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
            f0_mel_max - f0_mel_min
        ) + 1
        f0_mel[f0_mel <= 1] = 1
        f0_mel[f0_mel > 255] = 255
        f0_coarse = np.rint(f0_mel).astype(np.int)
        return f0_coarse, f0bak  # 1-0

    def vc(
        self,
        model,
        net_g,
        sid,
        audio0,
        pitch,
        pitchf,
        times,
        index,
        big_npy,
        index_rate,
    ):  # ,file_index,file_big_npy
        feats = torch.from_numpy(audio0)
        if self.is_half:
            feats = feats.half()
        else:
            feats = feats.float()
        if feats.dim() == 2:  # double channels
            feats = feats.mean(-1)
        assert feats.dim() == 1, feats.dim()
        feats = feats.view(1, -1)
        padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)

        inputs = {
            "source": feats.to(self.device),
            "padding_mask": padding_mask,
            "output_layer": 9,  # layer 9
        }
        t0 = ttime()
        with torch.no_grad():
            logits = model.extract_features(**inputs)
            feats = model.final_proj(logits[0])

        if (
            isinstance(index, type(None)) == False
            and isinstance(big_npy, type(None)) == False
            and index_rate != 0
        ):
            npy = feats[0].cpu().numpy()
            if self.is_half:
                npy = npy.astype("float32")

            # _, I = index.search(npy, 1)
            # npy = big_npy[I.squeeze()]

            # by github @nadare881
            score, ix = index.search(npy, k=8)
            weight = np.square(1 / score)
            weight /= weight.sum(axis=1, keepdims=True)
            npy = np.sum(big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)

            if self.is_half:
                npy = npy.astype("float16")
            feats = (
                torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate
                + (1 - index_rate) * feats
            )

        feats = F.interpolate(
            feats.permute(0, 2, 1),
            scale_factor=2,  # (index_rate) ** (-1)  # NOTE (Sam): why do we assume upsampling is 2x if input to Hubert is 16000 and default target rate is 40000?
        ).permute(0, 2, 1)
        t1 = ttime()
        p_len = audio0.shape[0] // self.window
        if feats.shape[1] < p_len:
            p_len = feats.shape[1]
            if pitch != None and pitchf != None:
                pitch = pitch[:, :p_len]
                pitchf = pitchf[:, :p_len]
        p_len = torch.tensor([p_len], device=self.device).long()
        with torch.no_grad():
            if pitch != None and pitchf != None:
                audio1 = (
                    (net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0] * 32768)
                    .data.cpu()
                    .float()
                    .numpy()
                    .astype(np.int16)
                )
            else:
                audio1 = (
                    (net_g.infer(feats, p_len, sid)[0][0, 0] * 32768)
                    .data.cpu()
                    .float()
                    .numpy()
                    .astype(np.int16)
                )
        del feats, p_len, padding_mask
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        t2 = ttime()
        times[0] += t1 - t0
        times[2] += t2 - t1
        return audio1