Support CharacterMix TimeLine

.gitignore

@@ -157,3 +157,10 @@ filelists/val.txt
 .idea/vcs.xml
 .idea/inspectionProfiles/profiles_settings.xml
 .idea/inspectionProfiles/Project_Default.xml
+pretrain/vec-768-layer-12.onnx
+pretrain/hubert-soft.onnx
+pretrain/hubert4.0.onnx
+pretrain/vec-256-layer-9.onnx
+pretrain/vec-256-layer-12.onnx
+pretrain/vec-768-layer-9.onnx
+.vscode/launch.json

inference/infer_tool.py

@@ -219,16 +219,23 @@ class Svc(object):
               f0_predictor='pm',
               enhancer_adaptive_key = 0,
               cr_threshold = 0.05,
-              k_step = 100
+              k_step = 100,
+              frame = 0,
+              spk_mix = False
               ):
-
-        speaker_id = self.spk2id.get(speaker)
-        if not speaker_id and type(speaker) is int:
-            if len(self.spk2id.__dict__) >= speaker:
-                speaker_id = speaker
-        sid = torch.LongTensor([int(speaker_id)]).to(self.dev).unsqueeze(0)
         wav, sr = librosa.load(raw_path, sr=self.target_sample)
-        c, f0, uv = self.get_unit_f0(wav, tran, cluster_infer_ratio, speaker, f0_filter,f0_predictor,cr_threshold=cr_threshold)
+        if spk_mix:
+            c, f0, uv = self.get_unit_f0(wav, tran, 0, None, f0_filter,f0_predictor,cr_threshold=cr_threshold)
+            n_frames = f0.size(1)
+            sid = speaker[:, frame:frame+n_frames].transpose(0,1)
+        else:
+            speaker_id = self.spk2id.get(speaker)
+            if not speaker_id and type(speaker) is int:
+                if len(self.spk2id.__dict__) >= speaker:
+                    speaker_id = speaker
+            sid = torch.LongTensor([int(speaker_id)]).to(self.dev).unsqueeze(0)
+            c, f0, uv = self.get_unit_f0(wav, tran, cluster_infer_ratio, speaker, f0_filter,f0_predictor,cr_threshold=cr_threshold)
+            n_frames = f0.size(1)
         if "half" in self.net_g_path and torch.cuda.is_available():
             c = c.half()
         with torch.no_grad():
@@ -266,7 +273,7 @@ class Svc(object):
                                     adaptive_key = enhancer_adaptive_key)
             use_time = time.time() - start
             print("vits use time:{}".format(use_time))
-        return audio, audio.shape[-1]
+        return audio, audio.shape[-1], n_frames
 
     def clear_empty(self):
         # clean up vram
@@ -297,7 +304,8 @@ class Svc(object):
                         f0_predictor='pm',
                         enhancer_adaptive_key = 0,
                         cr_threshold = 0.05,
-                        k_step = 100
+                        k_step = 100,
+                        use_spk_mix = False
                         ):
         wav_path = Path(raw_audio_path).with_suffix('.wav')
         chunks = slicer.cut(wav_path, db_thresh=slice_db)
@@ -308,7 +316,59 @@ class Svc(object):
         lg_size_c_l = (lg_size-lg_size_r)//2
         lg_size_c_r = lg_size-lg_size_r-lg_size_c_l
         lg = np.linspace(0,1,lg_size_r) if lg_size!=0 else 0
-        
+
+        if use_spk_mix:
+            assert len(self.spk2id) == len(spk)
+            self.net_g_ms.EnableCharacterMix(len(self.spk2id), self.dev)
+            audio_length = 0
+            for (slice_tag, data) in audio_data:
+                aud_length = int(np.ceil(len(data) / audio_sr * self.target_sample))
+                if slice_tag:
+                    audio_length += aud_length // self.hop_size
+                    continue
+                if per_size != 0:
+                    datas = split_list_by_n(data, per_size,lg_size)
+                else:
+                    datas = [data]
+                for k,dat in enumerate(datas):
+                    pad_len = int(audio_sr * pad_seconds)
+                    per_length = int(np.ceil(len(dat) / audio_sr * self.target_sample))
+                    a_length = per_length + 2 * pad_len
+                    audio_length += a_length // self.hop_size
+            audio_length += len(audio_data)
+            spk_mix_tensor = torch.zeros(size=(len(spk), audio_length)).to(self.dev)
+            for i in range(len(spk)):
+                last_end = None
+                for mix in spk[i]:
+                    if mix[3]<0. or mix[2]<0.:
+                        raise RuntimeError("mix value must higer Than zero!")
+                    begin = int(audio_length * mix[0])
+                    end = int(audio_length * mix[1])
+                    length = end - begin
+                    if length<=0:                        raise RuntimeError("begin Must lower Than end!")
+                    step = (mix[3] - mix[2])/length
+                    if last_end is not None:
+                        if last_end != begin:
+                            raise RuntimeError("[i]EndTime Must Equal [i+1]BeginTime!")
+                    last_end = end
+                    spk_mix_data = torch.arange(mix[2],mix[3],step).to(self.dev)
+                    if(len(spk_mix_data)<length):
+                        num_pad = length - len(spk_mix_data)
+                        spk_mix_data = torch.nn.functional.pad(spk_mix_data, [0, num_pad], mode="reflect").to(self.dev)
+                    spk_mix_tensor[i][begin:end] = spk_mix_data[:length]
+
+            spk_mix_ten = torch.sum(spk_mix_tensor,dim=0).unsqueeze(0).to(self.dev)
+            # spk_mix_tensor[0][spk_mix_ten<0.001] = 1.0
+            for i, x in enumerate(spk_mix_ten[0]):
+                if x == 0.0:
+                    spk_mix_ten[0][i] = 1.0
+                    spk_mix_tensor[:,i] = 1.0 / len(spk)
+            spk_mix_tensor = spk_mix_tensor / spk_mix_ten
+            if not ((torch.sum(spk_mix_tensor,dim=0) - 1.)<0.0001).all():
+                raise RuntimeError("sum(spk_mix_tensor) not equal 1")
+            spk = spk_mix_tensor
+
+        global_frame = 0
         audio = []
         for (slice_tag, data) in audio_data:
             print(f'#=====segment start, {round(len(data) / audio_sr, 3)}s======')
@@ -318,6 +378,7 @@ class Svc(object):
                 print('jump empty segment')
                 _audio = np.zeros(length)
                 audio.extend(list(pad_array(_audio, length)))
+                global_frame += length // self.hop_size
                 continue
             if per_size != 0:
                 datas = split_list_by_n(data, per_size,lg_size)
@@ -332,15 +393,18 @@ class Svc(object):
                 raw_path = io.BytesIO()
                 soundfile.write(raw_path, dat, audio_sr, format="wav")
                 raw_path.seek(0)
-                out_audio, out_sr = self.infer(spk, tran, raw_path,
+                out_audio, out_sr, out_frame = self.infer(spk, tran, raw_path,
                                                     cluster_infer_ratio=cluster_infer_ratio,
                                                     auto_predict_f0=auto_predict_f0,
                                                     noice_scale=noice_scale,
                                                     f0_predictor = f0_predictor,
                                                     enhancer_adaptive_key = enhancer_adaptive_key,
                                                     cr_threshold = cr_threshold,
-                                                    k_step = k_step
+                                                    k_step = k_step,
+                                                    frame = global_frame,
+                                                    spk_mix = use_spk_mix
                                                     )
+                global_frame += out_frame
                 _audio = out_audio.cpu().numpy()
                 pad_len = int(self.target_sample * pad_seconds)
                 _audio = _audio[pad_len:-pad_len]

inference_main.py

@@ -2,12 +2,11 @@ import io
 import logging
 import time
 from pathlib import Path
-
+from spkmix import spk_mix_map
 import librosa
 import matplotlib.pyplot as plt
 import numpy as np
 import soundfile
-
 from inference import infer_tool
 from inference import slicer
 from inference.infer_tool import Svc
@@ -23,10 +22,10 @@ def main():
     parser = argparse.ArgumentParser(description='sovits4 inference')
 
     # 一定要设置的部分
-    parser.add_argument('-m', '--model_path', type=str, default="logs/44k/G_0.pth", help='模型路径')
+    parser.add_argument('-m', '--model_path', type=str, default="logs/44k/G_37600.pth", help='模型路径')
-    parser.add_argument('-c', '--config_path', type=str, default="configs/config.json", help='配置文件路径')
+    parser.add_argument('-c', '--config_path', type=str, default="logs/44k/config.json", help='配置文件路径')
     parser.add_argument('-cl', '--clip', type=float, default=0, help='音频强制切片，默认0为自动切片，单位为秒/s')
-    parser.add_argument('-n', '--clean_names', type=str, nargs='+', default=["君の知らない物語-src.wav"], help='wav文件名列表，放在raw文件夹下')
+    parser.add_argument('-n', '--clean_names', type=str, nargs='+', default=["test.wav"], help='wav文件名列表，放在raw文件夹下')
     parser.add_argument('-t', '--trans', type=int, nargs='+', default=[0], help='音高调整，支持正负（半音）')
     parser.add_argument('-s', '--spk_list', type=str, nargs='+', default=['nen'], help='合成目标说话人名称')
     
@@ -38,6 +37,7 @@ def main():
     parser.add_argument('-f0p', '--f0_predictor', type=str, default="pm", help='选择F0预测器,可选择crepe,pm,dio,harvest,默认为pm(注意：crepe为原F0使用均值滤波器)')
     parser.add_argument('-eh', '--enhance', action='store_true', default=False, help='是否使用NSF_HIFIGAN增强器,该选项对部分训练集少的模型有一定的音质增强效果，但是对训练好的模型有反面效果，默认关闭')
     parser.add_argument('-shd', '--shallow_diffusion', action='store_true', default=False, help='是否使用浅层扩散，使用后可解决一部分电音问题，默认关闭，该选项打开时，NSF_HIFIGAN增强器将会被禁止')
+    parser.add_argument('-usm', '--use_spk_mix', action='store_true', default=True, help='是否使用角色融合')
 
     # 浅扩散设置
     parser.add_argument('-dm', '--diffusion_model_path', type=str, default="logs/44k/diffusion/model_0.pt", help='扩散模型路径')
@@ -79,9 +79,12 @@ def main():
     k_step = args.k_step
     only_diffusion = args.only_diffusion
     shallow_diffusion = args.shallow_diffusion
-
+    use_spk_mix = args.use_spk_mix
     svc_model = Svc(args.model_path, args.config_path, args.device, args.cluster_model_path,enhance,diffusion_model_path,diffusion_config_path,shallow_diffusion,only_diffusion)
     infer_tool.mkdir(["raw", "results"])
+
+    if use_spk_mix:
+        spk_list = [spk_mix_map]
     
     infer_tool.fill_a_to_b(trans, clean_names)
     for clean_name, tran in zip(clean_names, trans):
@@ -105,7 +108,8 @@ def main():
                 "f0_predictor" : f0p,
                 "enhancer_adaptive_key" : enhancer_adaptive_key,
                 "cr_threshold" : cr_threshold,
-                "k_step":k_step
+                "k_step":k_step,
+                "use_spk_mix":use_spk_mix
             }
             audio = svc_model.slice_inference(**kwarg)
             key = "auto" if auto_predict_f0 else f"{tran}key"
@@ -113,7 +117,9 @@ def main():
             isdiffusion = "sovits"
             if shallow_diffusion : isdiffusion = "sovdiff"
             if only_diffusion : isdiffusion = "diff"
-            res_path = f'./results/{clean_name}_{key}_{spk}{cluster_name}_{isdiffusion}.{wav_format}'
+            if type(spk) != type('aa'):
+                spk = "spk_mix"
+            res_path = f'results/{clean_name}_{key}_{spk}{cluster_name}_{isdiffusion}.{wav_format}'
             soundfile.write(res_path, audio, svc_model.target_sample, format=wav_format)
             svc_model.clear_empty()
             

models.py

@@ -382,11 +382,11 @@ class SynthesizerTrn(nn.Module):
         self.emb_uv = nn.Embedding(2, hidden_channels)
         self.character_mix = False
 
-    def EnableCharacterMix(self, n_speakers_map):
+    def EnableCharacterMix(self, n_speakers_map, device):
-        self.speaker_map = torch.zeros((n_speakers_map, 1, 1, self.gin_channels))
+        self.speaker_map = torch.zeros((n_speakers_map, 1, 1, self.gin_channels)).to(device)
         for i in range(n_speakers_map):
-            self.speaker_map[i] = self.emb_g(torch.LongTensor([[i]]))
+            self.speaker_map[i] = self.emb_g(torch.LongTensor([[i]]).to(device))
-        self.speaker_map = self.speaker_map.unsqueeze(0)
+        self.speaker_map = self.speaker_map.unsqueeze(0).to(device)
         self.character_mix = True
 
     def forward(self, c, f0, uv, spec, g=None, c_lengths=None, spec_lengths=None):
@@ -413,9 +413,25 @@ class SynthesizerTrn(nn.Module):
 
         return o, ids_slice, spec_mask, (z, z_p, m_p, logs_p, m_q, logs_q), pred_lf0, norm_lf0, lf0
 
-    def infer(self, c, f0, uv, g=None, noice_scale=0.35, predict_f0=False):
+    def infer(self, c, f0, uv, g=None, noice_scale=0.35, seed=52468, predict_f0=False):
+
+        if c.device == torch.device("cuda"):
+            torch.cuda.manual_seed_all(seed)
+        else:
+            torch.manual_seed(seed)
+
         c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device)
-        g = self.emb_g(g).transpose(1, 2)
+
+        if self.character_mix and len(g) > 1:   # [N, S]  *  [S, B, 1, H]
+            g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1))  # [N, S, B, 1, 1]
+            g = g * self.speaker_map  # [N, S, B, 1, H]
+            g = torch.sum(g, dim=1) # [N, 1, B, 1, H]
+            g = g.transpose(0, -1).transpose(0, -2).squeeze(0) # [B, H, N]
+        else:
+            if g.dim() == 1:
+                g = g.unsqueeze(0)
+            g = self.emb_g(g).transpose(1, 2)
+        
         x_mask = torch.unsqueeze(commons.sequence_mask(c_lengths, c.size(2)), 1).to(c.dtype)
         x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1, 2)
 

spkmix.py

@@ -0,0 +1,11 @@
+# 角色混合轨道 编写规则：
+# 角色ID : [[起始时间1, 终止时间1, 起始数值1, 起始数值1], [起始时间2, 终止时间2, 起始数值2, 起始数值2]]
+# 起始时间和前一个的终止时间必须相同，第一个起始时间必须为0，最后一个终止时间必须为1 （时间的范围为0-1）
+# 全部角色必须填写，不使用的角色填[[0., 1., 0., 0.]]即可
+# 融合数值可以随便填，在指定的时间段内从起始数值线性变化为终止数值，内部会自动确保线性组合为1，可以放心使用
+
+spk_mix_map = {
+    0 : [[0., 0.5, 1, 0.5], [0.5, 1, 0.5, 1]],
+    1 : [[0., 0.35, 1, 0.5], [0.35, 0.75, 0.75, 1], [0.75, 1, 0.45, 1]],
+    2 : [[0., 0.35, 1, 0.5], [0.35, 0.75, 0.75, 1], [0.75, 1, 0.45, 1]]
+}

utils.py

@@ -9,7 +9,6 @@ import subprocess
 import warnings
 import random
 import functools
-
 import librosa
 import numpy as np
 from scipy.io.wavfile import read
@@ -110,19 +109,19 @@ def get_speech_encoder(speech_encoder,device=None,**kargs):
         speech_encoder_object = ContentVec256L9(device = device)
     elif speech_encoder == "vec256l9-onnx":
         from vencoder.ContentVec256L9_Onnx import ContentVec256L9_Onnx
-        speech_encoder_object = ContentVec256L9(device = device)
+        speech_encoder_object = ContentVec256L9_Onnx(device = device)
     elif speech_encoder == "vec256l12-onnx":
         from vencoder.ContentVec256L12_Onnx import ContentVec256L12_Onnx
-        speech_encoder_object = ContentVec256L9(device = device)
+        speech_encoder_object = ContentVec256L12_Onnx(device = device)
     elif speech_encoder == "vec768l9-onnx":
         from vencoder.ContentVec768L9_Onnx import ContentVec768L9_Onnx
-        speech_encoder_object = ContentVec256L9(device = device)
+        speech_encoder_object = ContentVec768L9_Onnx(device = device)
     elif speech_encoder == "vec768l12-onnx":
         from vencoder.ContentVec768L12_Onnx import ContentVec768L12_Onnx
-        speech_encoder_object = ContentVec256L9(device = device)
+        speech_encoder_object = ContentVec768L12_Onnx(device = device)
     elif speech_encoder == "hubertsoft-onnx":
         from vencoder.HubertSoft_Onnx import HubertSoft_Onnx
-        speech_encoder_object = HubertSoft(device = device)
+        speech_encoder_object = HubertSoft_Onnx(device = device)
     elif speech_encoder == "hubertsoft":
         from vencoder.HubertSoft import HubertSoft
         speech_encoder_object = HubertSoft(device = device)