feat: add web ui and idiot-proof launch script

README.md

@@ -19,6 +19,12 @@ CosyVoice3 SFT fine-tune for a Mandarin-speaking talking flower character. Outpu
 
 This bundle is self-contained — no separate CosyVoice repository clone required.
 
+### Launching the Web UI
+To easily launch the interactive web interface, use the included shell script:
+```bash
+./launch-web-app.sh
+```
+
 ### Option 1: Using `uv` (Recommended)
 This is the fastest and most reliable way to run the model, leveraging a fully locked and isolated environment.
 

launch-web-app.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+# A foolproof script to launch the Talking Flower Web UI
+# It ensures dependencies are installed via `uv` and starts the Gradio app.
+
+set -e # Exit on error
+
+echo "🌸 Starting Talking Flower Web UI..."
+echo ""
+
+# Check if uv is installed
+if ! command -v uv &> /dev/null; then
+    echo "❌ Error: 'uv' is not installed."
+    echo "Please install it first: pip install uv"
+    exit 1
+fi
+
+# Change to the directory where this script is located
+cd "$(dirname "$0")"
+
+# Optional: Default environment variables, overridden if already set
+export GRADIO_SERVER_NAME="${GRADIO_SERVER_NAME:-0.0.0.0}"
+export GRADIO_SERVER_PORT="${GRADIO_SERVER_PORT:-6112}"
+
+echo "📦 Syncing dependencies using uv..."
+uv sync
+
+echo ""
+echo "🚀 Launching the Web Application..."
+echo "🌐 The UI will be available on all network interfaces (0.0.0.0) on port ${GRADIO_SERVER_PORT}"
+echo "Local access: http://localhost:${GRADIO_SERVER_PORT}"
+echo "Network access: http://<your-machine-ip>:${GRADIO_SERVER_PORT}"
+echo ""
+
+# Run the app
+uv run web-app/app.py

web-app/app.py

@@ -0,0 +1,174 @@
+import os
+import sys
+import time
+import random
+import torch
+import soundfile as sf
+import gradio as gr
+
+# Add the parent directory to sys.path so we can import cosyvoice and transformers
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_PARENT = os.path.dirname(_HERE)
+sys.path.insert(0, _PARENT)
+sys.path.insert(0, os.path.join(_PARENT, "third_party/Matcha-TTS"))
+
+import onnxruntime
+import transformers
+
+# Monkey patch same as inference.py to avoid missing files errors in HF slim-bundle
+_original_inference_session = onnxruntime.InferenceSession
+
+def _maybe_inference_session(path_or_bytes, *args, **kwargs):
+    if isinstance(path_or_bytes, str) and not os.path.exists(path_or_bytes):
+        return None
+    return _original_inference_session(path_or_bytes, *args, **kwargs)
+
+onnxruntime.InferenceSession = _maybe_inference_session
+
+_original_qwen2_from_pretrained = transformers.Qwen2ForCausalLM.from_pretrained
+
+def _qwen2_from_pretrained_or_config(pretrained_model_name_or_path, *args, **kwargs):
+    weight_files = ('model.safetensors', 'pytorch_model.bin',
+                    'model.safetensors.index.json', 'pytorch_model.bin.index.json')
+    if (isinstance(pretrained_model_name_or_path, str)
+            and os.path.isdir(pretrained_model_name_or_path)
+            and not any(os.path.exists(os.path.join(pretrained_model_name_or_path, f))
+                        for f in weight_files)):
+        config = transformers.Qwen2Config.from_pretrained(pretrained_model_name_or_path)
+        return transformers.Qwen2ForCausalLM(config)
+    return _original_qwen2_from_pretrained(pretrained_model_name_or_path, *args, **kwargs)
+
+transformers.Qwen2ForCausalLM.from_pretrained = _qwen2_from_pretrained_or_config
+
+from cosyvoice.cli.cosyvoice import CosyVoice3
+from cosyvoice.utils.common import set_all_random_seed
+
+MODEL_DIR = _PARENT
+INSTRUCT = "You are a helpful assistant.<|endofprompt|>"
+SPK_ID = "TalkingFlower"
+
+# Global model instance (lazy load)
+model = None
+
+def load_model():
+    global model
+    if model is None:
+        model = CosyVoice3(MODEL_DIR)
+    return model
+
+def remove_tail_click(audio, sr, search_s=0.20, burst_thresh=0.05,
+                      silence_thresh=0.02, win_ms=5, fade_ms=3):
+    ch = audio[0]
+    win_n = int(sr * win_ms / 1000)
+    search_n = min(int(sr * search_s), ch.shape[0])
+    fade_n = int(sr * fade_ms / 1000)
+    tail = ch[-search_n:]
+    n_wins = search_n // win_n
+    rms = [tail[i * win_n:(i + 1) * win_n].pow(2).mean().sqrt().item()
+           for i in range(n_wins)]
+    if rms[-1] < burst_thresh:
+        return audio
+    cut_win = next((i for i in range(n_wins - 2, -1, -1)
+                    if rms[i] < silence_thresh), None)
+    if cut_win is None:
+        return audio
+    cut = ch.shape[0] - search_n + cut_win * win_n
+    out = audio.clone()
+    out[0, cut:] = 0.0
+    if fade_n > 0 and cut >= fade_n:
+        out[0, cut - fade_n:cut] *= torch.linspace(1.0, 0.0, fade_n)
+    return out
+
+def generate_audio(text, seed, speed):
+    if not text:
+        return None, "Please enter some text."
+    
+    set_all_random_seed(seed)
+    
+    try:
+        model = load_model()
+        
+        for output in model.inference_sft(
+            INSTRUCT + text,
+            spk_id=SPK_ID,
+            stream=False,
+            speed=speed,
+            text_frontend=False,
+        ):
+            audio = remove_tail_click(output["tts_speech"], model.sample_rate)
+            
+            output_path = os.path.join(_HERE, "output.wav")
+            sf.write(output_path, audio.squeeze(0).cpu().numpy(), model.sample_rate)
+            return output_path, f"Success! (Seed: {seed})"
+            
+    except Exception as e:
+        return None, f"Error: {str(e)}"
+
+# Gradio UI Theme & Setup (inspired by Talking-Flower)
+custom_css = """
+#main-container { max-width: 900px; margin: auto; }
+.wonder-card { border-radius: 12px; box-shadow: 0 4px 6px rgba(0,0,0,0.05); }
+.header { text-align: center; margin-bottom: 2rem; }
+.model-arch { background-color: #f8f9fa; padding: 1rem; border-radius: 8px; border-left: 4px solid #f472b6; }
+"""
+
+with gr.Blocks(title="Talking Flower TTS", css=custom_css) as demo:
+    with gr.Column(elem_id="main-container"):
+        gr.Markdown("<div class='header'><h1>🌸 Talking Flower Web UI</h1><p>A CosyVoice3-powered TTS interface mimicking the original Talking-Flower.</p></div>")
+        
+        with gr.Row():
+            with gr.Column(scale=5):
+                text_input = gr.Textbox(
+                    label="Talking Flower will say:",
+                    lines=3,
+                    placeholder="Support Chinese, English and Japanese...",
+                    value="你好呀！我是会说话的花朵，很高兴认识你！",
+                    elem_classes="wonder-card"
+                )
+                
+                with gr.Row():
+                    seed_input = gr.Slider(
+                        label="Random Seed (For Reproducibility)",
+                        minimum=0, maximum=100000, step=1, value=42,
+                        elem_classes="wonder-card"
+                    )
+                    speed_input = gr.Slider(
+                        label="Speech Speed",
+                        minimum=0.5, maximum=2.0, step=0.1, value=1.0,
+                        elem_classes="wonder-card"
+                    )
+                    
+                generate_btn = gr.Button("🌸 Speak!", variant="primary", elem_classes="wonder-card")
+                
+            with gr.Column(scale=3):
+                audio_output = gr.Audio(
+                    label="输出音频 (Generated Audio)",
+                    type="filepath",
+                    interactive=False, # Enables native download button
+                    elem_classes="wonder-card"
+                )
+                status_output = gr.Textbox(label="Status", interactive=False, elem_classes="wonder-card")
+                
+        gr.Markdown("---")
+        gr.Markdown("### 🧠 Model Architecture (CosyVoice 3 Sub-Models)")
+        gr.HTML("""
+        <div class="model-arch">
+            <p>This repository uses a three-stage cascade architecture for zero-shot and supervised text-to-speech:</p>
+            <ol>
+                <li><strong>LLM (Language Model) <code>llm.pt</code></strong>: A Qwen2-0.5B based transformer that autoregressively generates semantic speech tokens from the input text and instruction.</li>
+                <li><strong>Flow Matching <code>flow.pt</code></strong>: A conditional flow-matching network that translates the discrete semantic tokens into continuous Mel-spectrogram features, conditioned on speaker embeddings.</li>
+                <li><strong>HIFT (HiFi-GAN Vocoder) <code>hift.pt</code></strong>: A high-fidelity generative adversarial network that converts the Mel-spectrograms into the final raw audio waveform.</li>
+            </ol>
+        </div>
+        """)
+
+        generate_btn.click(
+            fn=generate_audio,
+            inputs=[text_input, seed_input, speed_input],
+            outputs=[audio_output, status_output]
+        )
+
+if __name__ == "__main__":
+    port = int(os.environ.get("GRADIO_SERVER_PORT", 6112))
+    host = os.environ.get("GRADIO_SERVER_NAME", "0.0.0.0")
+    demo.launch(server_name=host, server_port=port)