Upload app.py

app.py

@@ -8,6 +8,7 @@ import os
 import json
 import re
 import tempfile
+import time
 import subprocess
 import shutil
 import struct
@@ -32,6 +33,10 @@ except ImportError:
 OMNI_MODEL = "qwen3.5-omni-plus"
 DASHSCOPE_BASE_URL = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
 
+# Persistent output directory (survives between requests)
+OUTPUT_DIR = os.path.join(tempfile.gettempdir(), "qwen3_tts_outputs")
+os.makedirs(OUTPUT_DIR, exist_ok=True)
+
 LANGUAGES = ["Auto", "English", "Chinese", "Japanese", "Korean", "German",
              "French", "Russian", "Portuguese", "Spanish", "Italian"]
 
@@ -120,7 +125,7 @@ def generate_custom_voice(text, language, speaker_label, instruction):
         kwargs["instruct"] = instruction.strip()
     print(f"[TTS] Custom: speaker={speaker}, lang={lang}")
     wavs, sr = model.generate_custom_voice(**kwargs)
-    path = os.path.join(tempfile.mkdtemp(), "custom.wav")
+    path = os.path.join(OUTPUT_DIR, f"custom_{int(time.time())}.wav")
     sf.write(path, wavs[0], sr)
     return path
 
@@ -135,7 +140,7 @@ def generate_voice_design(text, language, voice_description):
     lang = language if language != "Auto" else "Auto"
     print(f"[TTS] Design: lang={lang}, desc={voice_description[:60]}")
     wavs, sr = model.generate_voice_design(text=text, language=lang, instruct=voice_description)
-    path = os.path.join(tempfile.mkdtemp(), "design.wav")
+    path = os.path.join(OUTPUT_DIR, f"design_{int(time.time())}.wav")
     sf.write(path, wavs[0], sr)
     return path
 
@@ -155,7 +160,7 @@ def generate_voice_clone(text, language, ref_audio, ref_text):
         kwargs["x_vector_only_mode"] = True
     print(f"[TTS] Clone: lang={lang}")
     wavs, sr = model.generate_voice_clone(**kwargs)
-    path = os.path.join(tempfile.mkdtemp(), "clone.wav")
+    path = os.path.join(OUTPUT_DIR, f"clone_{int(time.time())}.wav")
     sf.write(path, wavs[0], sr)
     return path
 
@@ -299,7 +304,8 @@ def generate_multi_speaker_story(text_input, file_input, language, progress=gr.P
             "Add it in Settings > Secrets."
         )
 
-    tmp_dir = tempfile.mkdtemp(prefix="multispeaker_")
+    tmp_dir = os.path.join(OUTPUT_DIR, f"ms_{int(time.time())}")
+    os.makedirs(tmp_dir, exist_ok=True)
 
     # Step 1: Detect characters and emotions
     progress(0.05, desc="Analyzing story characters and emotions...")
@@ -439,16 +445,7 @@ The old man was silent for a long time. "I already did," he finally whispered. "
 
 DESCRIPTION = """
 # Qwen3-TTS Demo
-### Open-Source Text-to-Speech (1.7B) — Self-Hosted, No API Keys for TTS
-
-| Mode | What it does |
-|------|-------------|
-| **Custom Voice** | Pick a preset voice + optional emotion instruction |
-| **Voice Design** | Describe any voice in natural language |
-| **Voice Clone** | Clone a voice from 3 seconds of audio |
-| **Multi-Speaker Story** | Auto-detect characters, assign voices, add emotions — full audiobook |
 
-10 languages supported. Running on ZeroGPU (free).
 """
 
 with gr.Blocks(title="Qwen3-TTS Demo") as demo:
@@ -545,11 +542,9 @@ with gr.Blocks(title="Qwen3-TTS Demo") as demo:
 
     gr.Markdown(
         "---\n"
-        "**Models:** Qwen3-TTS-12Hz-1.7B (Apache 2.0) | "
         "**Languages:** EN, ZH, JA, KO, DE, FR, RU, PT, ES, IT | "
-        "**TTS:** Self-hosted on ZeroGPU | "
         "**Character Analysis:** Cloud AI (optional, for Multi-Speaker mode)"
     )
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch(allowed_paths=[OUTPUT_DIR])