Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+female_shadowheart4.flac filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,31 @@
 ---
-title: Chatterbox Flash Tts
-emoji: 🏃
-colorFrom: purple
-colorTo: blue
+title: Chatterbox-Flash TTS
+emoji: 🎙️
+colorFrom: red
+colorTo: indigo
 sdk: gradio
 sdk_version: 6.26.0
-python_version: '3.12'
 app_file: app.py
-pinned: false
+short_description: Block-diffusion zero-shot text-to-speech by Resemble AI
+python_version: "3.12"
+startup_duration_timeout: 30m
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Chatterbox-Flash TTS
+
+[Chatterbox-Flash](https://github.com/resemble-ai/chatterbox-flash) is a prior-calibrated block-diffusion zero-shot text-to-speech model by Resemble AI. It extends Chatterbox-TTS with a parallel masked decoder that enables streaming generation with significantly lower latency.
+
+## How it works
+
+1. Enter the text you want to synthesize.
+2. Optionally upload a short reference audio clip to clone a specific voice.
+3. Click **Generate** to produce speech.
+
+The model uses a block-diffusion decoder (based on Llama-520M) with prior-calibrated PMI scoring and early decoding via a time-shifted quantile schedule, paired with the original S3Gen flow-matching vocoder and GE2E voice encoder.
+
+## Reference
+
+- **Paper**: [Chatterbox-Flash: Prior-Calibrated Block Diffusion for Streaming Zero-Shot TTS](https://huggingface.co/papers/2605.30748)
+- **Model**: [ResembleAI/chatterbox-flash](https://huggingface.co/ResembleAI/chatterbox-flash)
+- **Code**: [github.com/resemble-ai/chatterbox-flash](https://github.com/resemble-ai/chatterbox-flash)
+- **License**: MIT

app.py

@@ -0,0 +1,160 @@
+import os
+os.environ.setdefault("CHATTERBOX_FLASH_ENGINE", "torch")
+
+import spaces
+import torch
+import numpy as np
+import gradio as gr
+
+# Install chatterbox-tts and chatterbox-flash with --no-deps to avoid
+# the torch==2.6.0 pin (incompatible with ZeroGPU's torch 2.11).
+# Then their real deps are listed in requirements.txt.
+import subprocess
+import sys
+
+subprocess.run(
+    [sys.executable, "-m", "pip", "install", "--no-deps",
+     "chatterbox-tts==0.1.7", "chatterbox-flash==0.1.0"],
+    check=True,
+)
+
+from chatterbox_flash import ChatterboxFlashTTS
+
+MODEL_ID = "ResembleAI/chatterbox-flash"
+DEVICE = "cuda"
+
+print(f"Loading Chatterbox-Flash from {MODEL_ID}...")
+tts = ChatterboxFlashTTS.from_pretrained(
+    MODEL_ID, device=DEVICE, dtype=torch.bfloat16,
+)
+tts.to(DEVICE)
+print("Model loaded successfully.")
+
+
+@spaces.GPU(duration=120)
+def generate_tts(
+    text_input: str,
+    audio_prompt_path: str | None = None,
+    exaggeration: float = 0.5,
+    temperature: float = 0.6,
+    cfg_scale: float = 1.0,
+    num_steps: int = 10,
+    seed_num: int = 0,
+):
+    """Generate speech from text using Chatterbox-Flash block-diffusion TTS.
+
+    Zero-shot voice cloning: provide a short reference audio clip to clone
+    the speaker's voice, or leave it empty to use the default voice.
+
+    Args:
+        text_input: Text to synthesize (max 300 characters).
+        audio_prompt_path: Reference audio file for voice cloning (optional).
+        exaggeration: Controls speech expressiveness (0.25-2.0, 0.5=neutral).
+        temperature: Controls randomness in generation (0.05-5.0, higher=more varied).
+        cfg_scale: Classifier-free guidance scale (0.2-1.0).
+        num_steps: Number of block-diffusion denoising steps (default 10).
+        seed_num: Random seed for reproducibility (0=random).
+    """
+    if seed_num != 0:
+        torch.manual_seed(int(seed_num))
+        torch.cuda.manual_seed(int(seed_num))
+        np.random.seed(int(seed_num))
+
+    generate_kwargs = {
+        "exaggeration": exaggeration,
+        "temperature": temperature,
+        "cfg_scale": cfg_scale,
+        "num_steps": num_steps,
+        "backend": "torch",
+    }
+
+    if audio_prompt_path:
+        generate_kwargs["audio_prompt_path"] = audio_prompt_path
+
+    wav = tts.generate(text_input[:300], **generate_kwargs)
+    return (tts.sr, wav.squeeze(0).numpy())
+
+
+CSS = """
+#col-container { max-width: 1100px; margin: 0 auto; }
+.dark .gradio-container { color: var(--body-text-color); }
+"""
+
+with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
+    gr.Markdown(
+        """
+        # Chatterbox-Flash TTS
+        Prior-calibrated block-diffusion zero-shot TTS by Resemble AI.
+        Provide a reference audio clip to clone a voice, or generate with the default voice.
+
+        [Paper](https://huggingface.co/papers/2605.30748) · [Model](https://huggingface.co/ResembleAI/chatterbox-flash) · [GitHub](https://github.com/resemble-ai/chatterbox-flash)
+        """
+    )
+
+    with gr.Row(id="col-container"):
+        with gr.Column(scale=3):
+            text = gr.Textbox(
+                value="Sometimes it's better to just let things slide, you know?",
+                label="Text to synthesize (max 300 chars)",
+                max_lines=5,
+            )
+            ref_wav = gr.Audio(
+                sources=["upload", "microphone"],
+                type="filepath",
+                label="Reference Audio (optional — for voice cloning)",
+            )
+
+            with gr.Accordion("Advanced settings", open=False):
+                exaggeration = gr.Slider(
+                    0.25, 2.0, step=0.05,
+                    label="Exaggeration (0.5=neutral, higher=more expressive)",
+                    value=0.5,
+                )
+                temperature = gr.Slider(
+                    0.05, 2.0, step=0.05,
+                    label="Temperature",
+                    value=0.6,
+                )
+                cfg_scale = gr.Slider(
+                    0.2, 1.0, step=0.05,
+                    label="CFG Scale",
+                    value=1.0,
+                )
+                num_steps = gr.Slider(
+                    1, 30, step=1,
+                    label="Denoising Steps",
+                    value=10,
+                )
+                seed_num = gr.Number(
+                    value=0, label="Seed (0=random)", precision=0,
+                )
+
+            run_btn = gr.Button("Generate", variant="primary")
+
+        with gr.Column(scale=2):
+            audio_output = gr.Audio(label="Output Audio")
+
+    gr.Examples(
+        examples=[
+            ["Sometimes it's better to just let things slide, you know?"],
+            ["The quick brown fox jumps over the lazy dog. Pack my box with five dozen liquor jugs."],
+            ["In the depths of winter, I finally learned that within me lay an invincible summer."],
+        ],
+        inputs=[text],
+        outputs=[audio_output],
+        fn=generate_tts,
+        cache_examples=True,
+        cache_mode="lazy",
+    )
+
+    run_btn.click(
+        fn=generate_tts,
+        inputs=[
+            text, ref_wav, exaggeration,
+            temperature, cfg_scale, num_steps, seed_num,
+        ],
+        outputs=[audio_output],
+        api_name="generate",
+    )
+
+demo.launch(mcp_server=True)

female_shadowheart4.flac

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19c9562517aca84931af549a7a3578ebb033024f9dbcda7f99881338cba6124c
+size 573883

requirements.txt

@@ -0,0 +1,18 @@
+chatterbox-tts==0.1.7
+chatterbox-flash==0.1.0
+transformers==5.2.0
+diffusers==0.29.0
+torchaudio
+librosa==0.11.0
+s3tokenizer
+resemble-perth>=1.0.0
+conformer==0.3.2
+safetensors==0.5.3
+omegaconf
+numpy<2.0.0
+soundfile
+scipy
+inflect>=7.0
+Unidecode>=1.3
+pyloudnorm
+peft