Upload 2 files

rvc/infer/infer.py

@@ -0,0 +1,257 @@
+import os
+import sys
+import torch
+import numpy as np
+import soundfile as sf
+from vc_infer_pipeline import VC
+from rvc.lib.utils import load_audio
+from rvc.lib.tools.split_audio import process_audio, merge_audio
+from fairseq import checkpoint_utils
+from rvc.lib.infer_pack.models import (
+    SynthesizerTrnMs256NSFsid,
+    SynthesizerTrnMs256NSFsid_nono,
+    SynthesizerTrnMs768NSFsid,
+    SynthesizerTrnMs768NSFsid_nono,
+)
+
+from rvc.configs.config import Config
+
+config = Config()
+
+torch.manual_seed(114514)
+hubert_model = None
+
+
+def load_hubert():
+    global hubert_model
+    models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
+        ["hubert_base.pt"],
+        suffix="",
+    )
+    hubert_model = models[0]
+    hubert_model = hubert_model.to(config.device)
+    if config.is_half:
+        hubert_model = hubert_model.half()
+    else:
+        hubert_model = hubert_model.float()
+    hubert_model.eval()
+
+
+def vc_single(
+    sid=0,
+    input_audio_path=None,
+    f0_up_key=None,
+    f0_file=None,
+    f0_method=None,
+    file_index=None,
+    index_rate=None,
+    resample_sr=0,
+    rms_mix_rate=1,
+    protect=0.33,
+    hop_length=None,
+    output_path=None,
+    split_audio=False,
+):
+    global tgt_sr, net_g, vc, hubert_model, version
+
+    if input_audio_path is None:
+        return "Please, load an audio!", None
+
+    f0_up_key = int(f0_up_key)
+    try:
+        audio = load_audio(input_audio_path, 16000)
+        audio_max = np.abs(audio).max() / 0.95
+
+        if audio_max > 1:
+            audio /= audio_max
+
+        if not hubert_model:
+            load_hubert()
+        if_f0 = cpt.get("f0", 1)
+
+        file_index = (
+            file_index.strip(" ")
+            .strip('"')
+            .strip("\n")
+            .strip('"')
+            .strip(" ")
+            .replace("trained", "added")
+        )
+        if tgt_sr != resample_sr >= 16000:
+            tgt_sr = resample_sr
+        if split_audio == "True":
+            result, new_dir_path = process_audio(input_audio_path)
+            if result == "Error":
+                return "Error with Split Audio", None
+            dir_path = new_dir_path.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+            if dir_path != "":
+                paths = [
+                    os.path.join(root, name)
+                    for root, _, files in os.walk(dir_path, topdown=False)
+                    for name in files
+                    if name.endswith(".wav") and root == dir_path
+                ]
+            try:
+                for path in paths:
+                    info, opt = vc_single(
+                        sid,
+                        path,
+                        f0_up_key,
+                        None,
+                        f0_method,
+                        file_index,
+                        index_rate,
+                        resample_sr,
+                        rms_mix_rate,
+                        protect,
+                        hop_length,
+                        path,
+                        False,
+                    )
+                    #new_dir_path
+            except Exception as error:
+                print(error)
+                return "Error", None
+            print("Finished processing segmented audio, now merging audio...")
+            merge_timestamps_file = os.path.join(os.path.dirname(new_dir_path), f"{os.path.basename(input_audio_path).split('.')[0]}_timestamps.txt")
+            tgt_sr, audio_opt  = merge_audio(merge_timestamps_file)
+      
+        else:
+            audio_opt = vc.pipeline(
+                hubert_model,
+                net_g,
+                sid,
+                audio,
+                input_audio_path,
+                f0_up_key,
+                f0_method,
+                file_index,
+                index_rate,
+                if_f0,
+                filter_radius,
+                tgt_sr,
+                resample_sr,
+                rms_mix_rate,
+                version,
+                protect,
+                hop_length,
+                f0_file=f0_file,
+            )
+        
+
+        if output_path is not None:
+            sf.write(output_path, audio_opt, tgt_sr, format="WAV")
+
+        return (tgt_sr, audio_opt)
+
+    except Exception as error:
+        print(error)
+
+
+def get_vc(weight_root, sid):
+    global n_spk, tgt_sr, net_g, vc, cpt, version
+    if sid == "" or sid == []:
+        global hubert_model
+        if hubert_model is not None:
+            print("clean_empty_cache")
+            del net_g, n_spk, vc, hubert_model, tgt_sr  # ,cpt
+            hubert_model = net_g = n_spk = vc = hubert_model = tgt_sr = None
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+            if_f0 = cpt.get("f0", 1)
+            version = cpt.get("version", "v1")
+            if version == "v1":
+                if if_f0 == 1:
+                    net_g = SynthesizerTrnMs256NSFsid(
+                        *cpt["config"], is_half=config.is_half
+                    )
+                else:
+                    net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
+            elif version == "v2":
+                if if_f0 == 1:
+                    net_g = SynthesizerTrnMs768NSFsid(
+                        *cpt["config"], is_half=config.is_half
+                    )
+                else:
+                    net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
+            del net_g, cpt
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+            cpt = None
+    person = weight_root
+    cpt = torch.load(person, map_location="cpu")
+    tgt_sr = cpt["config"][-1]
+    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+    if_f0 = cpt.get("f0", 1)
+
+    version = cpt.get("version", "v1")
+    if version == "v1":
+        if if_f0 == 1:
+            net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=config.is_half)
+        else:
+            net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
+    elif version == "v2":
+        if if_f0 == 1:
+            net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=config.is_half)
+        else:
+            net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
+    del net_g.enc_q
+    print(net_g.load_state_dict(cpt["weight"], strict=False))
+    net_g.eval().to(config.device)
+    if config.is_half:
+        net_g = net_g.half()
+    else:
+        net_g = net_g.float()
+    vc = VC(tgt_sr, config)
+    n_spk = cpt["config"][-3]
+
+
+f0up_key = sys.argv[1]
+filter_radius = sys.argv[2]
+index_rate = float(sys.argv[3])
+hop_length = sys.argv[4]
+f0method = sys.argv[5]
+
+audio_input_path = sys.argv[6]
+audio_output_path = sys.argv[7]
+
+model_path = sys.argv[8]
+index_path = sys.argv[9]
+split_audio = sys.argv[10]
+
+sid = f0up_key
+input_audio = audio_input_path
+f0_pitch = f0up_key
+f0_file = None
+f0_method = f0method
+file_index = index_path
+index_rate = index_rate
+output_file = audio_output_path
+split_audio = split_audio
+
+get_vc(model_path, 0)
+
+try:
+    result, audio_opt = vc_single(
+        sid=0,
+        input_audio_path=input_audio,
+        f0_up_key=f0_pitch,
+        f0_file=None,
+        f0_method=f0_method,
+        file_index=file_index,
+        index_rate=index_rate,
+        hop_length=hop_length,
+        output_path=output_file,
+        split_audio=split_audio
+    )
+
+    if os.path.exists(output_file) and os.path.getsize(output_file) > 0:
+        message = result
+    else:
+        message = result
+
+    print(f"Conversion completed. Output file: '{output_file}'")
+
+except Exception as error:
+    print(f"Voice conversion failed: {error}")

rvc/infer/vc_infer_pipeline.py

@@ -0,0 +1,492 @@
+import numpy as np, parselmouth, torch, pdb, sys, os
+from time import time as ttime
+import torch.nn.functional as F
+import torchcrepe
+from torch import Tensor
+import scipy.signal as signal
+import pyworld, os, faiss, librosa, torchcrepe
+from scipy import signal
+from functools import lru_cache
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
+
+input_audio_path2wav = {}
+
+
+@lru_cache
+def cache_harvest_f0(input_audio_path, fs, f0max, f0min, frame_period):
+    audio = input_audio_path2wav[input_audio_path]
+    f0, t = pyworld.harvest(
+        audio,
+        fs=fs,
+        f0_ceil=f0max,
+        f0_floor=f0min,
+        frame_period=frame_period,
+    )
+    f0 = pyworld.stonemask(audio, f0, t, fs)
+    return f0
+
+
+def change_rms(data1, sr1, data2, sr2, rate):
+    rms1 = librosa.feature.rms(y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2)
+    rms2 = librosa.feature.rms(y=data2, frame_length=sr2 // 2 * 2, hop_length=sr2 // 2)
+    rms1 = torch.from_numpy(rms1)
+    rms1 = F.interpolate(
+        rms1.unsqueeze(0), size=data2.shape[0], mode="linear"
+    ).squeeze()
+    rms2 = torch.from_numpy(rms2)
+    rms2 = F.interpolate(
+        rms2.unsqueeze(0), size=data2.shape[0], mode="linear"
+    ).squeeze()
+    rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-6)
+    data2 *= (
+        torch.pow(rms1, torch.tensor(1 - rate))
+        * torch.pow(rms2, torch.tensor(rate - 1))
+    ).numpy()
+    return data2
+
+
+class VC(object):
+    def __init__(self, tgt_sr, config):
+        self.x_pad, self.x_query, self.x_center, self.x_max, self.is_half = (
+            config.x_pad,
+            config.x_query,
+            config.x_center,
+            config.x_max,
+            config.is_half,
+        )
+        self.sr = 16000
+        self.window = 160
+        self.t_pad = self.sr * self.x_pad
+        self.t_pad_tgt = tgt_sr * self.x_pad
+        self.t_pad2 = self.t_pad * 2
+        self.t_query = self.sr * self.x_query
+        self.t_center = self.sr * self.x_center
+        self.t_max = self.sr * self.x_max
+        self.device = config.device
+
+    def get_optimal_torch_device(self, index: int = 0) -> torch.device:
+        if torch.cuda.is_available():
+            return torch.device(f"cuda:{index % torch.cuda.device_count()}")
+        elif torch.backends.mps.is_available():
+            return torch.device("mps")
+        return torch.device("cpu")
+
+    def get_f0_crepe_computation(
+        self,
+        x,
+        f0_min,
+        f0_max,
+        p_len,
+        hop_length=120,
+        model="full",
+    ):
+        x = x.astype(np.float32)
+        x /= np.quantile(np.abs(x), 0.999)
+        torch_device = self.get_optimal_torch_device()
+        audio = torch.from_numpy(x).to(torch_device, copy=True)
+        audio = torch.unsqueeze(audio, dim=0)
+        if audio.ndim == 2 and audio.shape[0] > 1:
+            audio = torch.mean(audio, dim=0, keepdim=True).detach()
+        audio = audio.detach()
+        print("Initiating prediction with a hop_length of: " + str(hop_length))
+        pitch: Tensor = torchcrepe.predict(
+            audio,
+            self.sr,
+            hop_length,
+            f0_min,
+            f0_max,
+            model,
+            batch_size=hop_length * 2,
+            device=torch_device,
+            pad=True,
+        )
+        p_len = p_len or x.shape[0] // hop_length
+        source = np.array(pitch.squeeze(0).cpu().float().numpy())
+        source[source < 0.001] = np.nan
+        target = np.interp(
+            np.arange(0, len(source) * p_len, len(source)) / p_len,
+            np.arange(0, len(source)),
+            source,
+        )
+        f0 = np.nan_to_num(target)
+        return f0
+
+    def get_f0_official_crepe_computation(
+        self,
+        x,
+        f0_min,
+        f0_max,
+        model="full",
+    ):
+        batch_size = 512
+        audio = torch.tensor(np.copy(x))[None].float()
+        f0, pd = torchcrepe.predict(
+            audio,
+            self.sr,
+            self.window,
+            f0_min,
+            f0_max,
+            model,
+            batch_size=batch_size,
+            device=self.device,
+            return_periodicity=True,
+        )
+        pd = torchcrepe.filter.median(pd, 3)
+        f0 = torchcrepe.filter.mean(f0, 3)
+        f0[pd < 0.1] = 0
+        f0 = f0[0].cpu().numpy()
+        return f0
+
+    def get_f0(
+        self,
+        input_audio_path,
+        x,
+        p_len,
+        f0_up_key,
+        f0_method,
+        filter_radius,
+        hop_length,
+        inp_f0=None,
+    ):
+        global input_audio_path2wav
+        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"
+                )
+        elif f0_method == "harvest":
+            input_audio_path2wav[input_audio_path] = x.astype(np.double)
+            f0 = cache_harvest_f0(input_audio_path, self.sr, f0_max, f0_min, 10)
+            if filter_radius > 2:
+                f0 = signal.medfilt(f0, 3)
+        elif f0_method == "dio":
+            f0, t = pyworld.dio(
+                x.astype(np.double),
+                fs=self.sr,
+                f0_ceil=f0_max,
+                f0_floor=f0_min,
+                frame_period=10,
+            )
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sr)
+            f0 = signal.medfilt(f0, 3)
+        elif f0_method == "crepe":
+            f0 = self.get_f0_crepe_computation(x, f0_min, f0_max, p_len, hop_length)
+        elif f0_method == "crepe-tiny":
+            f0 = self.get_f0_crepe_computation(
+                x, f0_min, f0_max, p_len, hop_length, "tiny"
+            )
+        elif f0_method == "rmvpe":
+            if hasattr(self, "model_rmvpe") == False:
+                from rvc.lib.rmvpe import RMVPE
+
+                self.model_rmvpe = RMVPE(
+                    "rmvpe.pt", is_half=self.is_half, device=self.device
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+
+        f0 *= pow(2, f0_up_key / 12)
+        tf0 = self.sr // self.window
+        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
+            ]
+        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
+
+    def vc(
+        self,
+        model,
+        net_g,
+        sid,
+        audio0,
+        pitch,
+        pitchf,
+        index,
+        big_npy,
+        index_rate,
+        version,
+        protect,
+    ):
+        feats = torch.from_numpy(audio0)
+        if self.is_half:
+            feats = feats.half()
+        else:
+            feats = feats.float()
+        if feats.dim() == 2:
+            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 if version == "v1" else 12,
+        }
+        t0 = ttime()
+        with torch.no_grad():
+            logits = model.extract_features(**inputs)
+            feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+        if protect < 0.5 and pitch != None and pitchf != None:
+            feats0 = feats.clone()
+        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")
+
+            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).permute(0, 2, 1)
+        if protect < 0.5 and pitch != None and pitchf != None:
+            feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).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]
+
+        if protect < 0.5 and pitch != None and pitchf != None:
+            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)
+        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])
+                    .data.cpu()
+                    .float()
+                    .numpy()
+                )
+            else:
+                audio1 = (
+                    (net_g.infer(feats, p_len, sid)[0][0, 0]).data.cpu().float().numpy()
+                )
+        del feats, p_len, padding_mask
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        t2 = ttime()
+        return audio1
+
+    def pipeline(
+        self,
+        model,
+        net_g,
+        sid,
+        audio,
+        input_audio_path,
+        f0_up_key,
+        f0_method,
+        file_index,
+        index_rate,
+        if_f0,
+        filter_radius,
+        tgt_sr,
+        resample_sr,
+        rms_mix_rate,
+        version,
+        protect,
+        hop_length,
+        f0_file=None,
+    ):
+        if file_index != "" and os.path.exists(file_index) == True and index_rate != 0:
+            try:
+                index = faiss.read_index(file_index)
+                big_npy = index.reconstruct_n(0, index.ntotal)
+            except Exception as error:
+                print(error)
+                index = big_npy = None
+        else:
+            index = big_npy = None
+        audio = signal.filtfilt(bh, ah, audio)
+        audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
+        opt_ts = []
+        if audio_pad.shape[0] > self.t_max:
+            audio_sum = np.zeros_like(audio)
+            for i in range(self.window):
+                audio_sum += audio_pad[i : i - self.window]
+            for t in range(self.t_center, audio.shape[0], self.t_center):
+                opt_ts.append(
+                    t
+                    - self.t_query
+                    + np.where(
+                        np.abs(audio_sum[t - self.t_query : t + self.t_query])
+                        == np.abs(audio_sum[t - self.t_query : t + self.t_query]).min()
+                    )[0][0]
+                )
+        s = 0
+        audio_opt = []
+        t = None
+        t1 = ttime()
+        audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
+        p_len = audio_pad.shape[0] // self.window
+        inp_f0 = None
+        if hasattr(f0_file, "name") == True:
+            try:
+                with open(f0_file.name, "r") as f:
+                    lines = f.read().strip("\n").split("\n")
+                inp_f0 = []
+                for line in lines:
+                    inp_f0.append([float(i) for i in line.split(",")])
+                inp_f0 = np.array(inp_f0, dtype="float32")
+            except Exception as error:
+                print(error)
+        sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
+        pitch, pitchf = None, None
+        if if_f0 == 1:
+            pitch, pitchf = self.get_f0(
+                input_audio_path,
+                audio_pad,
+                p_len,
+                f0_up_key,
+                f0_method,
+                filter_radius,
+                hop_length,
+                inp_f0,
+            )
+            pitch = pitch[:p_len]
+            pitchf = pitchf[:p_len]
+            if self.device == "mps":
+                pitchf = pitchf.astype(np.float32)
+            pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
+            pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
+        t2 = ttime()
+        for t in opt_ts:
+            t = t // self.window * self.window
+            if if_f0 == 1:
+                audio_opt.append(
+                    self.vc(
+                        model,
+                        net_g,
+                        sid,
+                        audio_pad[s : t + self.t_pad2 + self.window],
+                        pitch[:, s // self.window : (t + self.t_pad2) // self.window],
+                        pitchf[:, s // self.window : (t + self.t_pad2) // self.window],
+                        index,
+                        big_npy,
+                        index_rate,
+                        version,
+                        protect,
+                    )[self.t_pad_tgt : -self.t_pad_tgt]
+                )
+            else:
+                audio_opt.append(
+                    self.vc(
+                        model,
+                        net_g,
+                        sid,
+                        audio_pad[s : t + self.t_pad2 + self.window],
+                        None,
+                        None,
+                        index,
+                        big_npy,
+                        index_rate,
+                        version,
+                        protect,
+                    )[self.t_pad_tgt : -self.t_pad_tgt]
+                )
+            s = t
+        if if_f0 == 1:
+            audio_opt.append(
+                self.vc(
+                    model,
+                    net_g,
+                    sid,
+                    audio_pad[t:],
+                    pitch[:, t // self.window :] if t is not None else pitch,
+                    pitchf[:, t // self.window :] if t is not None else pitchf,
+                    index,
+                    big_npy,
+                    index_rate,
+                    version,
+                    protect,
+                )[self.t_pad_tgt : -self.t_pad_tgt]
+            )
+        else:
+            audio_opt.append(
+                self.vc(
+                    model,
+                    net_g,
+                    sid,
+                    audio_pad[t:],
+                    None,
+                    None,
+                    index,
+                    big_npy,
+                    index_rate,
+                    version,
+                    protect,
+                )[self.t_pad_tgt : -self.t_pad_tgt]
+            )
+        audio_opt = np.concatenate(audio_opt)
+        if rms_mix_rate != 1:
+            audio_opt = change_rms(audio, 16000, audio_opt, tgt_sr, rms_mix_rate)
+        if resample_sr >= 16000 and tgt_sr != resample_sr:
+            audio_opt = librosa.resample(
+                audio_opt, orig_sr=tgt_sr, target_sr=resample_sr
+            )
+        audio_max = np.abs(audio_opt).max() / 0.99
+        max_int16 = 32768
+        if audio_max > 1:
+            max_int16 /= audio_max
+        audio_opt = (audio_opt * max_int16).astype(np.int16)
+        del pitch, pitchf, sid
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        return audio_opt