feat: CosyVoice 한국어 엔진 통합 (Apache-2.0, GPU)

- vendor/CosyVoice: 패치된 CosyVoice 코드 + Matcha-TTS 벤더링 - cosyvoice_worker: 전용 venv에서 상주하는 합성 워커(FastAPI, 모델 1회 로드) - app/engines/cosyvoice_engine: HTTP 프록시 엔진(속도/피치) - Dockerfile: 전용 venv(torch cu128) + 모델 베이킹 + 빌드 워밍업 - entrypoint.sh: 워커(8001) + 메인앱(8788) 동시 기동
2026-08-23 23:16:24 +09:00
parent b92d2be2af
commit 9c1009d16d
93 changed files with 74398 additions and 6 deletions
--- a/vendor/CosyVoice/cosyvoice/hifigan/f0_predictor.py
+++ b/vendor/CosyVoice/cosyvoice/hifigan/f0_predictor.py
@@ -0,0 +1,103 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Kai Hu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+import torch.nn as nn
+try:
+    from torch.nn.utils.parametrizations import weight_norm
+except ImportError:
+    from torch.nn.utils import weight_norm
+from cosyvoice.transformer.convolution import CausalConv1d
+
+
+class ConvRNNF0Predictor(nn.Module):
+    def __init__(self,
+                 num_class: int = 1,
+                 in_channels: int = 80,
+                 cond_channels: int = 512
+                 ):
+        super().__init__()
+
+        self.num_class = num_class
+        self.condnet = nn.Sequential(
+            weight_norm(
+                nn.Conv1d(in_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+        )
+        self.classifier = nn.Linear(in_features=cond_channels, out_features=self.num_class)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.condnet(x)
+        x = x.transpose(1, 2)
+        return torch.abs(self.classifier(x).squeeze(-1))
+
+
+class CausalConvRNNF0Predictor(nn.Module):
+    def __init__(self,
+                 num_class: int = 1,
+                 in_channels: int = 80,
+                 cond_channels: int = 512
+                 ):
+        super().__init__()
+
+        self.num_class = num_class
+        self.condnet = nn.Sequential(
+            weight_norm(
+                CausalConv1d(in_channels, cond_channels, kernel_size=4, causal_type='right')
+            ),
+            nn.ELU(),
+            weight_norm(
+                CausalConv1d(cond_channels, cond_channels, kernel_size=3, causal_type='left')
+            ),
+            nn.ELU(),
+            weight_norm(
+                CausalConv1d(cond_channels, cond_channels, kernel_size=3, causal_type='left')
+            ),
+            nn.ELU(),
+            weight_norm(
+                CausalConv1d(cond_channels, cond_channels, kernel_size=3, causal_type='left')
+            ),
+            nn.ELU(),
+            weight_norm(
+                CausalConv1d(cond_channels, cond_channels, kernel_size=3, causal_type='left')
+            ),
+            nn.ELU(),
+        )
+        self.classifier = nn.Linear(in_features=cond_channels, out_features=self.num_class)
+
+    def forward(self, x: torch.Tensor, finalize: bool = True) -> torch.Tensor:
+        if finalize is True:
+            x = self.condnet[0](x)
+        else:
+            x = self.condnet[0](x[:, :, :-self.condnet[0].causal_padding], x[:, :, -self.condnet[0].causal_padding:])
+        for i in range(1, len(self.condnet)):
+            x = self.condnet[i](x)
+        x = x.transpose(1, 2)
+        return torch.abs(self.classifier(x).squeeze(-1))