Spaces:

---

ASLP-lab

/

VoiceSculptor

like 1

Running on Zero

Project description (VoiceSculptor)

VoiceSculptor is an open-source, instruction-following text-to-speech (Instruct TTS) system developed by the ASLP Lab and collaborators. It enables users to design a voice from natural-language descriptions (e.g., “calm late-night radio host, low pitch, slow pace, slightly husky”) and optionally apply fine-grained attribute control (gender, age, speaking rate, pitch, volume, emotion). The designed voice can then be used as a prompt waveform for CosyVoice2-based voice cloning, supporting downstream speech synthesis and interactive voice experiences.

The project includes an interactive Gradio-style workflow for “voice design → generate multiple candidates → select best outputs,” making it ideal for creators, researchers, and educators who need rapid iteration over speaking styles and expressive delivery. VoiceSculptor reports competitive results on the Chinese subset of InstructTTSEval (APS/DSD/RP tasks), indicating strong instruction-following capability compared with both open and commercial baselines.

Why we request a GPU grant: VoiceSculptor runs an LLM-based speech token generator (LLaSA-family) plus neural codec decoding and optional verification. These steps are compute-intensive and require GPU to provide reasonable latency and to support multiple concurrent users on a public Space. The grant will allow us to host a stable public demo so the community can reproduce results, test instruction-following behaviors, and build upon the Apache-2.0 licensed codebase.

Open-source & responsible use: VoiceSculptor is released under Apache-2.0 and includes a usage disclaimer discouraging impersonation, fraud, or malicious voice cloning.