change to cpu

app.py

@@ -61,7 +61,21 @@ os.environ["COQUI_TOS_AGREED"] = "1"
 ##print(m)
 #m = model_name
 
-xtts_model = TTS("tts_models/multilingual/multi-dataset/xtts_v2", gpu=False)
+#xtts_model = TTS("tts_models/multilingual/multi-dataset/xtts_v2", gpu=False)
+
+device = "cpu"
+model_name = "tts_models/multilingual/multi-dataset/xtts_v2"
+print("⏳Downloading model")
+ModelManager().download_model(model_name)
+model_path = os.path.join(
+    get_user_data_dir("tts"), model_name.replace("/", "--")
+)
+
+config = XttsConfig()
+config.load_json(os.path.join(model_path, "config.json"))
+xtts_model = Xtts.init_from_config(config)
+xtts_model.load_checkpoint(config, checkpoint_dir=model_path, eval=True)
+xtts_model.to(device)
 
 
 

requirements.txt

@@ -1,5 +1,5 @@
 # Preinstall requirements from TTS 
-TTS @ git+https://github.com/coqui-ai/TTS@v0.20.6
+TTS @ git+https://github.com/coqui-ai/TTS@fa28f99f1508b5b5366539b2149963edcb80ba62
 pydantic==1.10.13
 python-multipart==0.0.6
 typing-extensions>=4.8.0

utils.py

@@ -8,6 +8,7 @@ import textwrap
 import time
 import uuid
 import wave
+import torch
 
 import emoji
 import gradio as gr
@@ -81,8 +82,37 @@ def detect_language(prompt, xtts_supported_languages=None):
 
     return language
     
-def get_voice_streaming(prompt, language, chatbot_voice, xtts_model, suffix="0"):
-    gpt_cond_latent, speaker_embedding = get_latents(chatbot_voice, xtts_model) 
+def get_voice_streaming(self,prompt, language, chatbot_voice, xtts_model, suffix="0"):
+
+    speaker = {
+            "speaker_embedding": xtts_model.speaker_manager.speakers["Claribel Dervla"][
+                "speaker_embedding"
+            ]
+            .cpu()
+            .squeeze()
+            .half()
+            .tolist(),
+            "gpt_cond_latent": xtts_model.speaker_manager.speakers["Claribel Dervla"][
+                "gpt_cond_latent"
+            ]
+            .cpu()
+            .squeeze()
+            .half()
+            .tolist(),
+        }
+    
+    speaker_embedding = (
+            torch.tensor(speaker.get("speaker_embedding"))
+            .unsqueeze(0)
+            .unsqueeze(-1)
+        )
+    
+    gpt_cond_latent = (
+        torch.tensor(speaker.get("gpt_cond_latent"))
+        .reshape((-1, 1024))
+        .unsqueeze(0)
+    )
+    #gpt_cond_latent, speaker_embedding = get_latents(chatbot_voice, xtts_model) 
     try:
         t0 = time.time()
         chunks = xtts_model.inference_stream(