A 4.5M-parameter TTS model runs 70x real-time on CPU and cost just $10 to train
mrkn1
42 points
3 comments
November 18, 2026
Related Discussions
Found 5 related stories in 155.2ms across 13,334 title embeddings via pgvector HNSW
- How we made a text-to-speech model respond in sub-50 ms toebee · 129 pts · August 21, 2026 · 65% similar
- Show HN: Nano-Edge-TTS – 70x realtime, Lightweight Kokoro-based TTS model drmhse · 13 pts · November 15, 2026 · 65% similar
- Show HN: Nari Qwen3-TTS and Qwen3-ASR – High accuracy, low latency and cost toebee · 75 pts · September 14, 2026 · 58% similar
- New model BDH-CQ costs $0.007 per task 11x less than OpenAI Luna even w 80% off wordpad · 11 pts · August 13, 2026 · 51% similar
- Show HN: OfflineTTS — Free browser-based TTS & STT that runs locally twainyoung · 11 pts · July 19, 2026 · 51% similar
Discussion Highlights (2 comments)
ig-88ms
I was irritated by the childlike voice at first, but they acknowledge that they were aiming for a twenty something female voice and plan to build a more mature voice.
ilyaizen
Hi there. I made CopySpeech, a desktop app that reads copied text aloud. Copy the same text twice to hear it. The default voice is tiny-voca, a 4.5M-parameter model. In my tests, it starts speaking about 0.1 seconds after a warm start. I'm happy with how fast it is. Repo: https://github.com/ilyaizen/CopySpeech Site: https://copyspeech.ilyaizen.link/