Update app.py

app.py

@@ -1,13 +1,16 @@
 import os
-import uuid
-import time
 import torch
+import time
+import uuid
 import gradio as gr
-from melo.api import TTS
+from openvoice import se_extractor
 from openvoice.api import ToneColorConverter
 
-# Set temporary cache locations for Hugging Face Spaces
+# Set writable cache directory for torch
 os.environ["TORCH_HOME"] = "/tmp/torch"
+os.makedirs("/tmp/torch", exist_ok=True)
+
+# Environment fixes for HF Spaces
 os.environ["HF_HOME"] = "/tmp/huggingface"
 os.environ["HF_HUB_CACHE"] = "/tmp/huggingface"
 os.environ["TRANSFORMERS_CACHE"] = "/tmp/huggingface"
@@ -15,53 +18,48 @@ os.environ["MPLCONFIGDIR"] = "/tmp"
 os.environ["XDG_CACHE_HOME"] = "/tmp"
 os.environ["XDG_CONFIG_HOME"] = "/tmp"
 os.environ["NUMBA_DISABLE_CACHE"] = "1"
-os.makedirs("/tmp/torch", exist_ok=True)
 os.makedirs("/tmp/huggingface", exist_ok=True)
 os.makedirs("/tmp/flagged", exist_ok=True)
 
-# Output folder
+# Set model paths
+ckpt_converter = "checkpoints/converter/config.json"
 output_dir = "/tmp/outputs"
 os.makedirs(output_dir, exist_ok=True)
 
-# Initialize tone converter
-ckpt_converter = "checkpoints/converter/config.json"
+# Initialize OpenVoice converter
 tone_color_converter = ToneColorConverter(ckpt_converter)
 
-# Device setting
-device = "cuda" if torch.cuda.is_available() else "cpu"
+# Speaker embedding cache
+ref_speaker_embed = None
 
 def clone_and_speak(text, speaker_wav):
     if not speaker_wav:
         return "Please upload a reference .wav file."
 
-    base_name = f"output_{int(time.time())}_{uuid.uuid4().hex[:6]}"
-    tmp_melo_path = f"{output_dir}/{base_name}_tmp.wav"
-    final_output_path = f"{output_dir}/{base_name}_converted.wav"
-
-    # Use English speaker model
-    model = TTS(language="EN", device=device)
-    speaker_ids = model.hps.data.spk2id
-    default_speaker_id = next(iter(speaker_ids.values()))
-
-    # Generate base TTS voice
-    model.tts_to_file(text, default_speaker_id, tmp_melo_path)
+    # Generate a unique filename
+    timestamp = str(int(time.time()))
+    base_name = f"output_{timestamp}_{uuid.uuid4().hex[:6]}"
+    output_wav = os.path.join(output_dir, f"{base_name}.wav")
 
-    # Use speaker_wav as reference to extract style embedding
-    from openvoice import se_extractor
-    ref_se, _ = se_extractor.get_se(speaker_wav, tone_color_converter, vad=False)
+    # Extract style from uploaded speaker voice
+    global ref_speaker_embed
+    ref_speaker_embed = se_extractor.get_se(speaker_wav, tone_color_converter, vad=False)
 
-    # Run the tone conversion
+    # Generate speech using base model
     tone_color_converter.convert(
-        audio_src_path=tmp_melo_path,
-        src_se=ref_se,
-        tgt_se=ref_se,
-        output_path=final_output_path,
-        message="@HuggingFace",
+        text=text,
+        speaker_id="openvoice",
+        language="en",
+        ref_speaker=speaker_wav,
+        ref_embed=ref_speaker_embed,
+        output_path=output_wav,
+        top_k=10,
+        temperature=0.3
     )
 
-    return final_output_path
+    return output_wav
 
-# Gradio interface
+# Gradio interface (exposed as global `demo` for HF Spaces)
 gr.Interface(
     fn=clone_and_speak,
     inputs=[
@@ -69,7 +67,7 @@ gr.Interface(
         gr.Audio(type="filepath", label="Upload a Reference Voice (.wav)")
     ],
     outputs=gr.Audio(label="Synthesized Output"),
-    flagging_dir="/tmp/flagged",
-    title="Text to Voice using Melo TTS + OpenVoice",
-    description="Use Melo TTS for base synthesis and OpenVoice to apply a reference speaker's tone.",
+    flagging_dir="/tmp/flagged",  # safe temporary dir
+    title="Text to Voice using OpenVoice",
+    description="Clone any voice (English) and generate speech using OpenVoice on CPU.",
 ).launch()