Browser TTS hack: 43.7MB quantized model hits 115x real-time with word timestamps
blixt · x · 2026-11-02
blixt built a WebGPU runtime for Phonon 31M, an Apache-2.0 TTS model: converting large weight matrices to 8-bit integers shrinks the download to 43.7 MB, enabling local streaming 24 kHz English speech in Chrome/Safari (including iPhone), plus per-word and per-token timestamps the official package doesn't offer.
Versus Gradium's phonon-tts: 115× vs 31× real-time on M3 Max, 73× vs 21× on RTX 2080 Ti, 26× vs 13× on iPhone 17 Pro; first audio in 3 ms vs 15 ms, cold start 46 ms vs 373 ms. Whisper heard zero word errors across 110 clips (vs 0.22%), with identical speaker similarity and UTMOS scores. The trick, per the author: "you can just prompt it" plus weight quantization.
More from Infra
- Space datacenter heat is just engineering; cosmic rays are the real challenge — shashib · 2026-11-02
- Louisiana's $300B+ project pipeline includes Meta Hyperion, Amazon and Hut 8 AI data centers — MatthewChang · 2026-11-02
- AI data centers destroy 250,000 trees, undercutting 'paperless' eco promises — im_mansigupta · 2026-11-02
- "Someone must solve the cost of serving coworking agents" — hyhieu226 · 2026-11-02
- SmolChat v17: On-Device GGUF LLMs on Android Now Ships an OpenAI-Compatible Server — shubham0204_dev · 2026-11-02
- 8 Coding Agents × 250K Context Each, Concurrently, on Two RTX 3090s (~600+ tok/s) — o2beast · 2026-11-02