2026-10-20 · America/Los_Angeles · 社区动态 · #19

TTS Model running fully in the browser, on the CPU: 5.5× faster than PyTorch

I've been working on running models fully client-side, and the first one is done: **Ming-3.0-tts-0.5B** running entirely in the browser. No server, no API key, no GPU. After the model loads, the page makes zero network requests while it speaks. Try it: https://huggingface.co/spaces/fusorlabs/ming-wasm Code (MIT): https://github.com/fusor-rs/ming-wasm Under the hood it's an LLM pipeline. A Qwen2-0.5B decoder produces one conditioning vector per 80 ms of audio, a small diffusion transformer turns that into an audio latent (10 flow-matching steps with CFG), and a streaming vocoder turns latents…

热度 45.5 / 100；排名与评分保留该期记录。