Spaces:
Running on Zero
Running on Zero
| title: SemBridge — Speech & Singing | |
| emoji: 🎤 | |
| colorFrom: indigo | |
| colorTo: green | |
| sdk: gradio | |
| sdk_version: 6.30.0 | |
| app_file: app.py | |
| short_description: Unified TTS, singing, and mixed speech+song generation | |
| python_version: "3.10" | |
| startup_duration_timeout: 1h | |
| # SemBridge — Speech & Singing, one continuous voice | |
| Demo for [ASLP-lab/SemBridge](https://huggingface.co/ASLP-lab/SemBridge) — | |
| **Sem**antic **T**oken Anchoring for **B**ridging discrete semantic supervision and | |
| continuous acoustic generation. The unified model performs **text-to-speech**, | |
| **note-conditioned singing (SVS)**, and **mixed speech+singing** with one | |
| reference voice, in English and Chinese. | |
| - **Tab 🗣️ Text to Speech** — type a sentence, choose a packaged reference voice | |
| (or upload your own recording + transcript), and listen. | |
| - **Tab 🎵 Singing** — enter a score as `lyric|pitch|duration_seconds` lines and | |
| render it with a sung reference voice. | |
| - **Tab 🎭 Mixed** — run one of the paper's official speech → singing → speech | |
| sequences end to end. | |
| Everything runs the authors' own `sembridge` inference package (bundled in this | |
| Space) against the released checkpoint: 64-D 25 Hz continuous latents, 2-frame | |
| patches, BigVGAN-Large / native causal decoding, 44.1 kHz mono WAV output, capped | |
| at 500 patches (40 s) per request. | |
| The ten packaged voices and the singing scores are the exact inputs from the | |
| official [demo page](https://tiamojames.github.io/SemBridge_Demo/) — Seed-TTS | |
| speech references and GTSinger / Opencpop sung references. Those assets retain | |
| their original rights and terms. | |
| ## Technical notes | |
| - ZeroGPU (`zero-a10g`), bf16, ~817M-parameter generator + 740M acoustic codecs. | |
| - `@spaces.GPU(duration=…)` uses a callable estimator so each request reserves | |
| GPU time based on its expected length (flow steps × patch count) instead of a | |
| flat worst case. | |
| - Reference audio is downmixed and resampled to 44.1 kHz automatically; provide | |
| ~4–10 s of clean audio plus its exact transcript. |