Premium UI, quota-smart GPU duration, reference auto-clean, quota panel

1) Premium dark UI with hero header, advanced settings accordion and tips. 2) Per-request GPU duration estimation so visitors only reserve the quota they need. 3) Reference clips are auto-trimmed and normalised for closer cloning. 4) Live ZeroGPU quota panel decoded from the visitor's token plus session usage tracking. 5) Sample voice examples.

README.md

@@ -1,33 +1,33 @@
 ---
-title: IndexTTS-2 Gradio App
+title: Offex Voice Studio
 emoji: 🎙️
-colorFrom: blue
-colorTo: indigo
+colorFrom: indigo
+colorTo: pink
 sdk: gradio
 sdk_version: 5.45.0
 app_file: app.py
 pinned: false
 ---
 
-# IndexTTS-2 Gradio App
+# Offex Voice Studio
 
-This Space provides a Gradio interface for English speech synthesis with IndexTTS-2.
+Premium Gradio interface for high-fidelity voice cloning with IndexTTS-2.
 
-## Build and runtime dependencies
+## Quota-smart design
 
-The Space SDK is pinned to Gradio 5.45.0. That version keeps huggingface-hub on the 0.x line required by the official IndexTTS 4.52.1 stack and avoids the Gradio 6.28/tokenizers resolver conflict.
+- All pip installs and the full model snapshot (main + auxiliary models, ~8 GB) are downloaded by a background warm-up thread on the CPU at Space startup, so the ZeroGPU allocation is used only for synthesis.
+- The GPU duration is estimated per request from the input length instead of a fixed value, so visitors only reserve the quota they actually need.
 
-The static requirements file contains the official IndexTTS runtime dependencies, including the CUDA 12.8 PyTorch 2.8 pins. It intentionally omits gradio, spaces, transformers, and tokenizers:
+## Quality features
 
-- Gradio and spaces are supplied by the Hugging Face Space runtime.
-- tokenizers==0.21.0 is installed lazily with --no-deps.
-- transformers==4.52.1 is installed lazily with --no-deps.
-- the official IndexTTS repository is installed lazily with --no-deps.
+- Reference clips are automatically trimmed of silence, capped to 20 seconds and peak-normalised for a closer clone.
+- Optional emotion reference and text-derived emotion controls.
+- Seed control for reproducible generations.
 
-The model snapshot is downloaded lazily to /tmp/index-tts-2 on the first generation request. The app uses IndexTTS2 from indextts.infer_v2 and passes the CUDA/configuration options required by IndexTTS-2.
+## Runtime notes
 
-## UI behavior
+The Space SDK is pinned to Gradio 5.45.0 to stay on the huggingface-hub 0.x line required by the official IndexTTS 4.52.1 stack.
 
-Upload or record a reference voice, enter English text, confirm the reference-voice permission checkbox, and press Generate Voice. An optional emotion reference, emotion strength, text-emotion checkbox, and seed are available.
+The static requirements file intentionally omits gradio, spaces, transformers and tokenizers; those are supplied by the Space runtime or installed lazily with --no-deps. tensorboard is pinned to 2.18.0 for NumPy 2.x compatibility (2.9.1 crashed on the removed `np.bool8` alias), and app.py also re-adds the alias defensively for any other legacy dependency.
 
-The launcher uses the standard Python guard and queues up to four generation requests.
+Upload or record a reference voice, enter English text, confirm the permission checkbox, and press Generate Voice.

app.py

@@ -5,6 +5,7 @@ import subprocess
 import sys
 import tempfile
 import threading
+import time
 from pathlib import Path
 from typing import Optional
 
@@ -86,6 +87,18 @@ def _download_model() -> Path:
     return MODEL_DIR
 
 
+def _download_aux_models() -> None:
+    """Pre-download auxiliary models (w2v-bert-2.0, MaskGCT codec, CAMPPlus,
+    BigVGAN) into the model cache on the CPU.
+
+    IndexTTS2 downloads these itself when given aux_paths=None, which would
+    happen inside the ZeroGPU allocation and waste quota on ~5 GB of downloads.
+    """
+    from indextts.utils.model_download import ensure_models_available
+
+    ensure_models_available(str(MODEL_DIR))
+
+
 def _patch_numpy_compat() -> None:
     """Restore NumPy aliases removed in NumPy 2.x for legacy pinned libraries.
 
@@ -98,16 +111,18 @@ def _patch_numpy_compat() -> None:
         np.bool8 = np.bool_
 
 
-def _download_aux_models() -> None:
-    """Pre-download auxiliary models (w2v-bert-2.0, MaskGCT codec, CAMPPlus,
-    BigVGAN) into the model cache on the CPU.
-
-    IndexTTS2 downloads these itself when given aux_paths=None, which would
-    happen inside the ZeroGPU allocation and waste quota on ~5 GB of downloads.
-    """
-    from indextts.utils.model_download import ensure_models_available
+def _download_example_voice() -> Optional[str]:
+    """Fetch one sample reference voice so users can try the app instantly."""
+    try:
+        from huggingface_hub import hf_hub_download
 
-    ensure_models_available(str(MODEL_DIR))
+        return hf_hub_download(
+            repo_id="IndexTeam/IndexTTS-2-Demo",
+            repo_type="space",
+            filename="examples/voice_02.wav",
+        )
+    except Exception:
+        return None
 
 
 def _warm_up() -> None:
@@ -133,6 +148,8 @@ def _warm_up() -> None:
 _warmup_thread = threading.Thread(target=_warm_up, daemon=True)
 _warmup_thread.start()
 
+EXAMPLE_VOICE = _download_example_voice()
+
 
 def _find_config(model_dir: Path) -> str:
     candidates = [
@@ -204,6 +221,29 @@ def _validate_audio(path: Optional[str], label: str) -> str:
     return str(audio)
 
 
+def _clean_reference(path: str) -> str:
+    """Trim silence, cap length and normalise the reference clip for a closer clone."""
+    try:
+        import librosa
+        import numpy as np
+        import soundfile as sf
+
+        audio, sample_rate = librosa.load(path, sr=None, mono=True)
+        trimmed, _ = librosa.effects.trim(audio, top_db=30)
+        if len(trimmed) < sample_rate // 2:
+            trimmed = audio
+        trimmed = trimmed[: int(sample_rate * 20.0)]
+        peak = float(np.max(np.abs(trimmed))) if trimmed.size else 0.0
+        if peak > 0:
+            trimmed = trimmed * (0.9 / peak)
+        cleaned = f"{path}.cleaned.wav"
+        sf.write(cleaned, trimmed, sample_rate)
+        return cleaned
+    except Exception as exc:
+        print(f"[reference-clean] skipped: {type(exc).__name__}: {exc}", flush=True)
+        return path
+
+
 def _set_seed(seed: int) -> None:
     random.seed(seed)
 
@@ -232,7 +272,29 @@ def _supported_kwargs(function, values: dict) -> dict:
     return {key: value for key, value in values.items() if key in parameters}
 
 
-@spaces.GPU(duration=150)
+def _estimate_duration(
+    reference_audio: Optional[str],
+    text: str,
+    permission: bool,
+    emotion_audio: Optional[str],
+    emotion_strength: float,
+    use_text_emotion: bool,
+    seed,
+) -> int:
+    """Request only as much ZeroGPU time as this request realistically needs.
+
+    The platform pre-checks the declared duration against the visitor's
+    remaining quota, so smaller realistic durations let people fit far more
+    generations into their daily allowance.
+    """
+    if _tts is None:
+        return 150  # cold start: one-time model load plus synthesis
+
+    characters = len((text or "").strip())
+    return int(min(120, 30 + characters * 0.05))
+
+
+@spaces.GPU(duration=_estimate_duration)
 def synthesize(
     reference_audio: Optional[str],
     text: str,
@@ -245,7 +307,7 @@ def synthesize(
     if not permission:
         raise gr.Error("Please confirm that you have permission to use the reference voice.")
 
-    reference_audio = _validate_audio(reference_audio, "Reference audio")
+    reference_audio = _clean_reference(_validate_audio(reference_audio, "Reference audio"))
     text = (text or "").strip()
 
     if not text:
@@ -272,7 +334,9 @@ def synthesize(
     if not isinstance(use_text_emotion, bool):
         raise gr.Error("Use text emotion must be a boolean.")
 
-    emotion_audio = _validate_audio(emotion_audio, "Emotion audio") if emotion_audio else None
+    emotion_audio = (
+        _clean_reference(_validate_audio(emotion_audio, "Emotion audio")) if emotion_audio else None
+    )
     output_dir = OUTPUT_DIR
     output_dir.mkdir(parents=True, exist_ok=True)
     output_path = output_dir / "index-tts-2-output.wav"
@@ -280,6 +344,7 @@ def synthesize(
     if output_path.exists():
         output_path.unlink()
 
+    started = time.perf_counter()
     try:
         _set_seed(numeric_seed)
         tts = _get_tts()
@@ -312,63 +377,179 @@ def synthesize(
         if not output_path.is_file() or output_path.stat().st_size == 0:
             raise RuntimeError("IndexTTS-2 returned without creating an audio file.")
 
-        return str(output_path)
+        elapsed = round(time.perf_counter() - started, 1)
+        return str(output_path), elapsed
     except gr.Error:
         raise
     except Exception as exc:
         raise gr.Error(f"Synthesis failed: {type(exc).__name__}: {exc}") from exc
 
 
-with gr.Blocks(title="Offex IndexTTS-2 Voice Clone") as demo:
-    gr.Markdown(
-        "# Offex IndexTTS-2 Voice Clone\n"
-        "Generate English speech from a permitted reference voice."
-    )
+# --------------------------------------------------------------------------
+# Quota panel helpers (CPU only, never touches the GPU)
+# --------------------------------------------------------------------------
 
-    with gr.Row():
-        reference_audio = gr.Audio(
-            label="Reference voice",
-            sources=["upload", "microphone"],
-            type="filepath",
-        )
-        emotion_audio = gr.Audio(
-            label="Optional emotion reference",
-            sources=["upload", "microphone"],
-            type="filepath",
-        )
+DAILY_HINT = "Free accounts get ~5 min/day, PRO ~40 min/day. The window resets 24h after your first use."
 
-    text = gr.Textbox(
-        label="English text",
-        placeholder="Enter the English text to synthesize...",
-        lines=6,
-    )
 
-    permission = gr.Checkbox(
-        label="I have permission to use this reference voice",
-        value=False,
-    )
+def _read_ip_token(request) -> dict:
+    """Decode the ZeroGPU identity token HF attaches to every request."""
+    import base64
+    import json
+
+    headers = getattr(request, "headers", None) or {}
+    token = headers.get("x-ip-token") or headers.get("X-IP-Token")
+    if not token:
+        return {}
+    try:
+        payload = json.loads(base64.urlsafe_b64decode(f"{token.split('.')[1]}=="))
+        return payload if isinstance(payload, dict) else {}
+    except Exception:
+        return {}
+
+
+def _quota_html(request, used_seconds: float) -> str:
+    claims = _read_ip_token(request)
+    identity = claims.get("user") or ("signed-in visitor" if claims.get("sub") else "guest (not signed in)")
+
+    used = max(0.0, float(used_seconds or 0))
+    bar = min(100, round(used / 300 * 100))  # free-tier daily allowance estimate
+    color = "#22c55e" if bar < 50 else "#eab308" if bar < 80 else "#ef4444"
+
+    return f"""
+    <div class="quota-card">
+      <div class="quota-title">⚡ ZeroGPU quota</div>
+      <div class="quota-row"><span>Identity</span><b>{identity}</b></div>
+      <div class="quota-row"><span>GPU time used on this Space (this session)</span><b>~{used:.0f}s</b></div>
+      <div class="quota-bar"><div class="quota-fill" style="width:{bar}%;background:{color};"></div></div>
+      <div class="quota-note">Bar = your session usage vs a ~5 min free daily allowance (estimate). {DAILY_HINT}</div>
+      <div class="quota-note">💡 This Space auto-trims requests to the minimum GPU time needed, so your quota lasts longer. Upgrade to <a href="https://huggingface.co/pricing" target="_blank" style="color:#818cf8;">HF PRO</a> for ~8x quota.</div>
+    </div>
+    """
+
+
+def _update_session_usage(used_seconds, last_elapsed, request: gr.Request):
+    total = float(used_seconds or 0) + float(last_elapsed or 0)
+    return total, _quota_html(request, total)
+
+
+def _refresh_quota(request: gr.Request):
+    return _quota_html(request, 0)
+
+
+# --------------------------------------------------------------------------
+# Premium UI
+# --------------------------------------------------------------------------
+
+CUSTOM_CSS = """
+.gradio-container {max-width: 1180px !important; margin: 0 auto !important;}
+#hero {
+  background: linear-gradient(120deg, #4f46e5 0%, #7c3aed 45%, #db2777 100%);
+  border-radius: 18px; padding: 26px 30px; margin-bottom: 8px;
+  box-shadow: 0 12px 32px rgba(79, 70, 229, 0.35);
+}
+#hero h1 {color: #ffffff; font-size: 2rem; font-weight: 800; margin: 0 0 6px 0; letter-spacing: -0.02em;}
+#hero p {color: rgba(255, 255, 255, 0.88); margin: 0; font-size: 0.98rem;}
+#hero .hero-badges {margin-top: 12px;}
+#hero .badge {
+  display: inline-block; background: rgba(255, 255, 255, 0.16); color: #fff;
+  border: 1px solid rgba(255, 255, 255, 0.25); border-radius: 999px;
+  padding: 3px 12px; font-size: 0.78rem; margin-right: 6px;
+}
+#generate-btn {
+  background: linear-gradient(120deg, #4f46e5, #db2777) !important;
+  border: none !important; color: #fff !important; font-weight: 700 !important;
+  font-size: 1.02rem !important; height: 52px !important;
+  box-shadow: 0 8px 22px rgba(219, 39, 119, 0.45) !important;
+}
+#generate-btn:hover {filter: brightness(1.08);}
+.quota-card {
+  background: #0f172a; border: 1px solid #1e293b; border-radius: 14px;
+  padding: 16px 20px; margin-top: 10px;
+}
+.quota-title {font-weight: 700; font-size: 1rem; margin-bottom: 10px; color: #e2e8f0;}
+.quota-row {display: flex; justify-content: space-between; font-size: 0.9rem; margin: 4px 0; color: #94a3b8;}
+.quota-row b {color: #e2e8f0;}
+.quota-bar {height: 10px; background: #1e293b; border-radius: 999px; overflow: hidden; margin: 10px 0;}
+.quota-fill {height: 100%; border-radius: 999px; transition: width .4s ease;}
+.quota-note {font-size: 0.78rem; color: #64748b; margin-top: 6px; line-height: 1.5;}
+.tips-card {font-size: 0.92rem; line-height: 1.7;}
+footer {display: none !important;}
+"""
+
+HEADER_HTML = """
+<div id="hero">
+  <h1>🎙️ Offex Voice Studio</h1>
+  <p>Clone any voice from a short reference clip — high fidelity, GPU-accelerated, powered by IndexTTS-2.</p>
+  <div class="hero-badges">
+    <span class="badge">🎧 Voice cloning</span>
+    <span class="badge">🎭 Emotion control</span>
+    <span class="badge">⚡ Quota-smart</span>
+  </div>
+</div>
+"""
+
+TIPS_MD = """
+**For the closest possible clone:**
+1. **Reference clip:** use 5–15 seconds of **clean speech** from one speaker only (no music, no background noise). The app auto-trims silence and normalises volume, but it cannot remove background noise.
+2. **Emotion reference:** add a second clip of the same voice speaking with the emotion you want — this dramatically improves expressiveness.
+3. **Keep texts short** — shorter texts synthesize faster and use less of your daily GPU quota.
+4. **Seed:** keep the same seed to get consistent output across retries; change it to explore variations.
+"""
+
+with gr.Blocks(title="Offex Voice Studio — AI Voice Cloning", theme=gr.themes.Soft(primary_hue="indigo", neutral_hue="slate"), css=CUSTOM_CSS) as demo:
+    gr.HTML(HEADER_HTML)
 
     with gr.Row():
-        emotion_strength = gr.Slider(
-            minimum=0,
-            maximum=1,
-            value=0.6,
-            step=0.05,
-            label="Emotion strength",
-        )
-        use_text_emotion = gr.Checkbox(
-            label="Use emotion information from text",
-            value=False,
-        )
-        seed = gr.Number(
-            label="Seed",
-            value=1234,
-            precision=0,
-            minimum=0,
-        )
+        with gr.Column(scale=5):
+            reference_audio = gr.Audio(
+                label="🎧 Reference voice (5–15s of clean speech works best)",
+                sources=["upload", "microphone"],
+                type="filepath",
+            )
+            with gr.Accordion("🎭 Emotion reference (optional — big quality boost)", open=False):
+                emotion_audio = gr.Audio(
+                    label="Emotion reference (same voice, desired emotion)",
+                    sources=["upload", "microphone"],
+                    type="filepath",
+                )
+        with gr.Column(scale=7):
+            text = gr.Textbox(
+                label="📝 Text to speak (English)",
+                placeholder="Type the text you want in the cloned voice...",
+                lines=5,
+                max_lines=12,
+            )
+            permission = gr.Checkbox(
+                label="✅ I have permission to use this reference voice",
+                value=False,
+            )
+            generate = gr.Button("✨ Generate Voice", variant="primary", size="lg", elem_id="generate-btn")
+            output = gr.Audio(label="🔊 Generated voice", type="filepath", autoplay=True, show_download_button=True)
+
+    with gr.Accordion("⚙️ Advanced settings", open=False):
+        with gr.Row():
+            emotion_strength = gr.Slider(
+                minimum=0,
+                maximum=1,
+                value=0.6,
+                step=0.05,
+                label="Emotion strength",
+            )
+            use_text_emotion = gr.Checkbox(
+                label="Derive emotion from the text itself",
+                value=False,
+            )
+            seed = gr.Number(
+                label="Seed (same seed = consistent output)",
+                value=1234,
+                precision=0,
+                minimum=0,
+            )
 
-    generate = gr.Button("Generate Voice", variant="primary", size="lg")
-    output = gr.Audio(label="Generated voice", type="filepath", autoplay=True)
+    usage_state = gr.State(0.0)
+    elapsed_box = gr.Number(visible=False, value=0)
+    quota_html = gr.HTML(elem_id="quota-panel")
 
     generate.click(
         fn=synthesize,
@@ -381,9 +562,28 @@ with gr.Blocks(title="Offex IndexTTS-2 Voice Clone") as demo:
             use_text_emotion,
             seed,
         ],
-        outputs=output,
+        outputs=[output, elapsed_box],
+    ).then(
+        fn=_update_session_usage,
+        inputs=[usage_state, elapsed_box],
+        outputs=[usage_state, quota_html],
     )
 
+    demo.load(fn=_refresh_quota, inputs=None, outputs=quota_html)
+
+    with gr.Accordion("💡 Tips for the most accurate clone", open=False):
+        gr.Markdown(TIPS_MD)
+
+    if EXAMPLE_VOICE:
+        gr.Examples(
+            examples=[
+                [EXAMPLE_VOICE, "Hello there! This is my cloned voice, speaking naturally and clearly."],
+                [EXAMPLE_VOICE, "Welcome to Offex Voice Studio. Today is a wonderful day to create something amazing."],
+            ],
+            inputs=[reference_audio, text],
+            label="Try it instantly with a sample voice",
+        )
+
 
 if __name__ == "__main__":
-    demo.queue(max_size=4).launch()
+    demo.queue(max_size=8).launch()