Upload 3 files

README.md

@@ -1,13 +1,37 @@
 ---
-title: Qwen3-TTS AudioBook Demo
-emoji: 🏢
-colorFrom: yellow
-colorTo: blue
+title: Qwen3-TTS Demo
+emoji: "\U0001F399"
+colorFrom: blue
+colorTo: purple
 sdk: gradio
-sdk_version: 6.17.3
-python_version: '3.13'
+sdk_version: 5.25.0
 app_file: app.py
 pinned: false
+license: apache-2.0
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Qwen3-TTS Demo
+
+Open-source text-to-speech with three modes:
+
+1. **Custom Voice** - Pick a preset speaker with optional emotion instructions
+2. **Voice Design** - Describe any voice in natural language and the AI creates it
+3. **Voice Clone** - Clone a voice from a 3-second audio sample
+
+## Setup
+
+No API keys needed. The models load automatically from HuggingFace.
+
+Hardware: Requires GPU (runs on ZeroGPU for free on HF Spaces).
+
+## Supported Languages
+
+English, Chinese, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian
+
+## Models Used
+
+- Qwen3-TTS-12Hz-1.7B-CustomVoice (preset voices)
+- Qwen3-TTS-12Hz-1.7B-VoiceDesign (natural language voice design)
+- Qwen3-TTS-12Hz-1.7B-Base (voice cloning)
+
+All models are Apache 2.0 licensed.

app.py

@@ -0,0 +1,252 @@
+"""
+Qwen3-TTS Demo — Self-hosted Text-to-Speech
+Three modes: Custom Voice, Voice Design, Voice Clone
+Runs on HF Spaces with ZeroGPU (free)
+"""
+
+import os
+import tempfile
+import torch
+import spaces
+import gradio as gr
+import soundfile as sf
+
+from qwen_tts import Qwen3TTSModel
+
+# ==========================================
+# MODEL LOADING
+# ==========================================
+# Models are loaded on-demand per mode to save VRAM
+_models = {}
+
+def get_model(model_type):
+    """Load model lazily. Models share the tokenizer so memory is manageable."""
+    if model_type not in _models:
+        model_map = {
+            "custom": "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
+            "design": "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
+            "clone": "Qwen/Qwen3-TTS-12Hz-1.7B-Base",
+        }
+        print(f"[TTS] Loading {model_map[model_type]}...")
+        _models[model_type] = Qwen3TTSModel.from_pretrained(
+            model_map[model_type],
+            device_map="cuda:0",
+            dtype=torch.bfloat16,
+        )
+        print(f"[TTS] {model_type} model loaded.")
+    return _models[model_type]
+
+
+# ==========================================
+# CONFIG
+# ==========================================
+LANGUAGES = ["Auto", "English", "Chinese", "Japanese", "Korean", "German",
+             "French", "Russian", "Portuguese", "Spanish", "Italian"]
+
+SPEAKERS = {
+    "Vivian": "Bright, edgy young female (Chinese)",
+    "Serena": "Warm, gentle young female (Chinese)",
+    "Uncle_Fu": "Seasoned male, low mellow timbre (Chinese)",
+    "Dylan": "Youthful Beijing male, clear natural (Chinese)",
+    "Eric": "Lively Chengdu male, slightly husky (Chinese/Sichuan)",
+    "Ryan": "Dynamic male, strong rhythmic drive (English)",
+    "Aiden": "Sunny American male, clear midrange (English)",
+    "Ono_Anna": "Playful Japanese female, light nimble (Japanese)",
+    "Sohee": "Warm Korean female, rich emotion (Korean)",
+}
+
+SPEAKER_CHOICES = [f"{name} -- {desc}" for name, desc in SPEAKERS.items()]
+
+EMOTION_EXAMPLES = [
+    "Very happy and excited",
+    "Speak sadly, with a heavy heart",
+    "Whisper softly and mysteriously",
+    "Angry and frustrated tone",
+    "Calm, warm bedtime story narrator",
+    "Professional news anchor delivery",
+    "Dramatic storytelling with suspense",
+]
+
+
+# ==========================================
+# TTS FUNCTIONS
+# ==========================================
+@spaces.GPU
+def generate_custom_voice(text, language, speaker_label, instruction):
+    """Mode 1: Custom Voice — pick a preset speaker with optional instruction."""
+    if not text.strip():
+        raise gr.Error("Please enter some text.")
+
+    model = get_model("custom")
+    speaker = speaker_label.split("--")[0].strip()
+    lang = language if language != "Auto" else "Auto"
+
+    kwargs = {
+        "text": text,
+        "language": lang,
+        "speaker": speaker,
+    }
+    if instruction and instruction.strip():
+        kwargs["instruct"] = instruction.strip()
+
+    print(f"[TTS] Custom voice: speaker={speaker}, lang={lang}, instruct={instruction[:50] if instruction else 'none'}")
+    wavs, sr = model.generate_custom_voice(**kwargs)
+
+    output_path = os.path.join(tempfile.mkdtemp(), "custom_voice.wav")
+    sf.write(output_path, wavs[0], sr)
+    print(f"[TTS] Generated: {output_path}, {len(wavs[0])/sr:.1f}s")
+    return output_path
+
+
+@spaces.GPU
+def generate_voice_design(text, language, voice_description):
+    """Mode 2: Voice Design — describe the voice you want in natural language."""
+    if not text.strip():
+        raise gr.Error("Please enter some text.")
+    if not voice_description.strip():
+        raise gr.Error("Please describe the voice you want.")
+
+    model = get_model("design")
+    lang = language if language != "Auto" else "Auto"
+
+    print(f"[TTS] Voice design: lang={lang}, desc={voice_description[:80]}")
+    wavs, sr = model.generate_voice_design(
+        text=text,
+        language=lang,
+        instruct=voice_description,
+    )
+
+    output_path = os.path.join(tempfile.mkdtemp(), "voice_design.wav")
+    sf.write(output_path, wavs[0], sr)
+    print(f"[TTS] Generated: {output_path}, {len(wavs[0])/sr:.1f}s")
+    return output_path
+
+
+@spaces.GPU
+def generate_voice_clone(text, language, ref_audio, ref_text):
+    """Mode 3: Voice Clone — clone a voice from a 3+ second audio sample."""
+    if not text.strip():
+        raise gr.Error("Please enter some text.")
+    if ref_audio is None:
+        raise gr.Error("Please upload a reference audio sample.")
+
+    model = get_model("clone")
+    lang = language if language != "Auto" else "Auto"
+
+    kwargs = {
+        "text": text,
+        "language": lang,
+        "ref_audio": ref_audio,
+    }
+    if ref_text and ref_text.strip():
+        kwargs["ref_text"] = ref_text.strip()
+    else:
+        kwargs["x_vector_only_mode"] = True
+
+    print(f"[TTS] Voice clone: lang={lang}, ref_text={'yes' if ref_text else 'speaker-embed only'}")
+    wavs, sr = model.generate_voice_clone(**kwargs)
+
+    output_path = os.path.join(tempfile.mkdtemp(), "voice_clone.wav")
+    sf.write(output_path, wavs[0], sr)
+    print(f"[TTS] Generated: {output_path}, {len(wavs[0])/sr:.1f}s")
+    return output_path
+
+
+# ==========================================
+# GRADIO UI
+# ==========================================
+DESCRIPTION = """
+# Qwen3-TTS Demo
+### Open-Source Text-to-Speech (1.7B)
+
+Three modes for generating speech:
+
+| Mode | What it does |
+|------|-------------|
+| **Custom Voice** | Pick a preset voice + optional emotion/style instruction |
+| **Voice Design** | Describe the voice you want in plain English |
+| **Voice Clone** | Clone any voice from a 3-second audio sample |
+
+Supports 10 languages: English, Chinese, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian.
+Running on ZeroGPU — completely free.
+"""
+
+with gr.Blocks(title="Qwen3-TTS Demo") as demo:
+
+    gr.Markdown(DESCRIPTION)
+
+    with gr.Tab("Custom Voice"):
+        gr.Markdown("Pick a preset speaker and optionally add emotion/style instructions.")
+        with gr.Row():
+            with gr.Column():
+                cv_text = gr.Textbox(label="Text to Speak", lines=4,
+                                     placeholder="Enter the text you want spoken...")
+                cv_lang = gr.Dropdown(choices=LANGUAGES, value="Auto", label="Language")
+                cv_speaker = gr.Dropdown(choices=SPEAKER_CHOICES,
+                                         value="Ryan -- Dynamic male, strong rhythmic drive (English)",
+                                         label="Speaker")
+                cv_instruct = gr.Textbox(label="Emotion / Style Instruction (optional)",
+                                         placeholder="e.g. Very happy and excited, Speak sadly, Whisper softly...")
+                cv_btn = gr.Button("Generate", variant="primary")
+            with gr.Column():
+                cv_audio = gr.Audio(label="Generated Speech", type="filepath")
+
+        cv_btn.click(fn=generate_custom_voice,
+                     inputs=[cv_text, cv_lang, cv_speaker, cv_instruct],
+                     outputs=cv_audio)
+
+    with gr.Tab("Voice Design"):
+        gr.Markdown("Describe the voice you want in natural language — the AI creates it from scratch.")
+        with gr.Row():
+            with gr.Column():
+                vd_text = gr.Textbox(label="Text to Speak", lines=4,
+                                     placeholder="Enter the text you want spoken...")
+                vd_lang = gr.Dropdown(choices=LANGUAGES, value="English", label="Language")
+                vd_desc = gr.Textbox(label="Voice Description", lines=3,
+                                     placeholder="e.g. Warm, captivating storyteller with a slight British accent, male...")
+                gr.Examples(
+                    examples=[[ex] for ex in [
+                        "Warm, captivating male storyteller with a slight British accent",
+                        "Young energetic female voice, cheerful and bright, American",
+                        "Deep authoritative male voice, news anchor style, very clear",
+                        "Gentle elderly grandmother voice, kind and soothing",
+                        "Speak in an incredulous tone with panic creeping into the voice",
+                    ]],
+                    inputs=[vd_desc],
+                    label="Example Descriptions",
+                )
+                vd_btn = gr.Button("Generate", variant="primary")
+            with gr.Column():
+                vd_audio = gr.Audio(label="Generated Speech", type="filepath")
+
+        vd_btn.click(fn=generate_voice_design,
+                     inputs=[vd_text, vd_lang, vd_desc],
+                     outputs=vd_audio)
+
+    with gr.Tab("Voice Clone"):
+        gr.Markdown("Clone any voice from a short audio sample (3+ seconds). Provide the transcript for best quality.")
+        with gr.Row():
+            with gr.Column():
+                vc_text = gr.Textbox(label="Text to Speak (in the cloned voice)", lines=4,
+                                     placeholder="Enter what you want the cloned voice to say...")
+                vc_lang = gr.Dropdown(choices=LANGUAGES, value="English", label="Language")
+                vc_ref_audio = gr.Audio(label="Reference Audio (3+ seconds)", type="filepath")
+                vc_ref_text = gr.Textbox(label="Transcript of Reference Audio (optional, improves quality)",
+                                         placeholder="Type what the person says in the reference audio...")
+                vc_btn = gr.Button("Clone & Generate", variant="primary")
+            with gr.Column():
+                vc_audio = gr.Audio(label="Generated Speech (Cloned Voice)", type="filepath")
+
+        vc_btn.click(fn=generate_voice_clone,
+                     inputs=[vc_text, vc_lang, vc_ref_audio, vc_ref_text],
+                     outputs=vc_audio)
+
+    gr.Markdown(
+        "---\n"
+        "**Model:** Qwen3-TTS-12Hz-1.7B (Apache 2.0) | "
+        "**Languages:** EN, ZH, JA, KO, DE, FR, RU, PT, ES, IT | "
+        "**Running on:** HF Spaces ZeroGPU"
+    )
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,5 @@
+qwen-tts>=0.1.0
+torch>=2.1.0
+soundfile>=0.12.0
+gradio>=5.25.0
+audioop-lts; python_version >= "3.13"