新增gpu模式下chattts代码

Chat2TTS/core.py

@@ -12,6 +12,9 @@ from .utils.io_utils import get_latest_modified_file
 from .infer.api import refine_text, infer_code
 from dataclasses import dataclass
 from typing import Literal, Optional, List, Tuple, Dict
+import numpy as np
+import pybase16384 as b14
+import lzma
 
 from huggingface_hub import snapshot_download
 
@@ -167,5 +170,23 @@ class Chat:
         
         return wav
         
-        
 
+    def sample_random_speaker(self) -> str:
+        return self._encode_spk_emb(self._sample_random_speaker())
+
+
+    @staticmethod
+    def _encode_spk_emb(spk_emb: torch.Tensor) -> str:
+        with torch.no_grad():
+            arr: np.ndarray = spk_emb.to(dtype=torch.float16, device="cpu").numpy()
+            s = b14.encode_to_string(
+                lzma.compress(
+                    arr.tobytes(),
+                    format=lzma.FORMAT_RAW,
+                    filters=[
+                        {"id": lzma.FILTER_LZMA2, "preset": 9 | lzma.PRESET_EXTREME}
+                    ],
+                ),
+            )
+            del arr
+        return s

web/app_gpu.py

@@ -29,7 +29,7 @@ def init_chat(args):
     source = "custom"
     # 获取启动模式
     MODEL = os.getenv('MODEL')
-    logger.info("loading ChatTTS model..., start MODEL:" + str(MODEL))
+    logger.info("loading Chat2TTS model..., start MODEL:" + str(MODEL))
     # huggingface 部署模式下，模型则直接使用hf的模型数据
     if MODEL == "HF":
         source = "huggingface"
@@ -253,25 +253,27 @@ def get_chat_infer_text(text,seed,refine_text_checkBox):
 def on_audio_seed_change(audio_seed_input):
     global chat
     torch.manual_seed(audio_seed_input)
-    rand_spk = torch.randn(audio_seed_input)
-    return encode_spk_emb(rand_spk)
-
-def encode_spk_emb(spk_emb: torch.Tensor) -> str:
-    import pybase16384 as b14
-    import lzma
-    with torch.no_grad():
-        arr: np.ndarray = spk_emb.to(dtype=torch.float16, device="cpu").numpy()
-        s = b14.encode_to_string(
-            lzma.compress(
-                arr.tobytes(),
-                format=lzma.FORMAT_RAW,
-                filters=[
-                    {"id": lzma.FILTER_LZMA2, "preset": 9 | lzma.PRESET_EXTREME}
-                ],
-            ),
-        )
-        del arr
-    return s
+    rand_spk = chat.sample_random_speaker()
+    return rand_spk
+    # rand_spk = torch.randn(audio_seed_input)
+    # return encode_spk_emb(rand_spk)
+
+# def encode_spk_emb(spk_emb: torch.Tensor) -> str:
+#     import pybase16384 as b14
+#     import lzma
+#     with torch.no_grad():
+#         arr: np.ndarray = spk_emb.to(dtype=torch.float16, device="cpu").numpy()
+#         s = b14.encode_to_string(
+#             lzma.compress(
+#                 arr.tobytes(),
+#                 format=lzma.FORMAT_RAW,
+#                 filters=[
+#                     {"id": lzma.FILTER_LZMA2, "preset": 9 | lzma.PRESET_EXTREME}
+#                 ],
+#             ),
+#         )
+#         del arr
+#     return s
 
 
 # def _sample_random_speaker(self) -> torch.Tensor: