Merge remote-tracking branch 'origin/main' into tts2dev

.github/workflows/docker-publish.yml

@@ -0,0 +1,44 @@
+name: Build and Publish Docker Image
+on:
+  workflow_dispatch:
+
+jobs:
+  build-amd64:
+    runs-on: ubuntu-22.04
+    strategy:
+      matrix:
+        include:
+          - cuda_version: 11.8
+            torch_version: 2.4.1
+            tag_prefix: pytorch2.4.1-cuda11.8
+          - cuda_version: 12.8
+            torch_version: 2.8.0
+            tag_prefix: pytorch2.8.0-cuda12.8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Extract Docker Meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: nanaoto/index-tts
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Build Docker Image
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./Dockerfile
+          push: false
+          platforms: linux/amd64
+          build-args: |
+            CUDA_VERSION=${{ matrix.cuda_version }}
+            TORCH_VERSION=${{ matrix.torch_version }}
+          tags: |
+            nanaoto/index-tts:${{ matrix.tag_prefix }}-${{ steps.meta.outputs.tags }}-amd64
+            nanaoto/index-tts:latest-${{ matrix.tag_prefix }}-amd64
+            
+            

indextts/infer_v2.py

@@ -100,12 +100,12 @@ class IndexTTS2:
         if self.use_cuda_kernel:
             # preload the CUDA kernel for BigVGAN
             try:
-                from indextts.BigVGAN.alias_free_activation.cuda import load
+                from indextts.s2mel.modules.bigvgan.alias_free_activation.cuda import activation1d
 
-                anti_alias_activation_cuda = load.load()
-                print(">> Preload custom CUDA kernel for BigVGAN", anti_alias_activation_cuda)
-            except:
+                print(">> Preload custom CUDA kernel for BigVGAN", activation1d.anti_alias_activation_cuda)
+            except Exception as e:
                 print(">> Failed to load custom CUDA kernel for BigVGAN. Falling back to torch.")
+                print(f"{e!r}")
                 self.use_cuda_kernel = False
 
         self.extract_features = SeamlessM4TFeatureExtractor.from_pretrained("facebook/w2v-bert-2.0")
@@ -292,6 +292,20 @@ class IndexTTS2:
         if self.gr_progress is not None:
             self.gr_progress(value, desc=desc)
 
+    def _load_and_cut_audio(self,audio_path,max_audio_length_seconds,verbose=False,sr=None):
+        if not sr:
+            audio, sr = librosa.load(audio_path)
+        else:
+            audio, _ = librosa.load(audio_path,sr=sr)
+        audio = torch.tensor(audio).unsqueeze(0)
+        max_audio_samples = int(max_audio_length_seconds * sr)
+
+        if audio.shape[1] > max_audio_samples:
+            if verbose:
+                print(f"Audio too long ({audio.shape[1]} samples), truncating to {max_audio_samples} samples")
+            audio = audio[:, :max_audio_samples]
+        return audio, sr
+
     # 原始推理模式
     def infer(self, spk_audio_prompt, text, output_path,
               emo_audio_prompt=None, emo_alpha=1.0,
@@ -340,8 +354,7 @@ class IndexTTS2:
 
         # 如果参考音频改变了，才需要重新生成, 提升速度
         if self.cache_spk_cond is None or self.cache_spk_audio_prompt != spk_audio_prompt:
-            audio, sr = librosa.load(spk_audio_prompt)
-            audio = torch.tensor(audio).unsqueeze(0)
+            audio,sr = self._load_and_cut_audio(spk_audio_prompt,15,verbose)
             audio_22k = torchaudio.transforms.Resample(sr, 22050)(audio)
             audio_16k = torchaudio.transforms.Resample(sr, 16000)(audio)
 
@@ -392,7 +405,7 @@ class IndexTTS2:
             emovec_mat = emovec_mat.unsqueeze(0)
 
         if self.cache_emo_cond is None or self.cache_emo_audio_prompt != emo_audio_prompt:
-            emo_audio, _ = librosa.load(emo_audio_prompt, sr=16000)
+            emo_audio, _ = self._load_and_cut_audio(emo_audio_prompt,15,verbose,sr=16000)
             emo_inputs = self.extract_features(emo_audio, sampling_rate=16000, return_tensors="pt")
             emo_input_features = emo_inputs["input_features"]
             emo_attention_mask = emo_inputs["attention_mask"]