NERDDISCO releases kale-tts-1, a 14.8MB open TTS family
NERDDISCO has released its first text-to-speech model family, kale-tts-1, on Hugging Face, with open weights in three sizes—base, tiny, and nano—along with the training dataset, training scripts, and the full recipe, plus an online demo. The release continues a recent wave of open-source TTS models that run on pure CPU.
Confirmed
- The nano version has just 7.5M parameters and a 14.8MB download footprint on pure CPU (30.6MB under WebGPU), with 10 built-in voices
- Performance: first audio within 90ms on pure CPU; under WebGPU, up to 125x real-time with 25ms time-to-first-sound; another figure cited is 44x real-time
- Architecture consists of a 0.9M dictionary guesser, a 3.8M text/timing planner, and a 3.7M third component (the generator part)
- Weights, training scripts, dataset, and full recipe are all open, with an online demo available
- Next step: integration into projects such as mlx-audio and llama.cpp
Why it matters
In a landscape where voice models often run several GB, kale-tts-1-nano shows that ultra-compact models can still deliver low-latency, high-quality on-device speech synthesis—fully open-source and reproducible—making it directly useful for voice applications in browsers and on edge devices (e.g., local assistants and real-time interaction).
2026-11-20 ~ 2026-11-21 · 7 related posts
Primary sources
- [source] kale-tts-1-nano: open-source 14.8MB TTS, 125x real-time on WebGPU — NERDDISCO · 2026-11-20
- [source] Kale TTS nano: a 14.8MB on-device TTS with 90ms first sound and 44x realtime — NERDDISCO · 2026-11-20
- How to extend Kale TTS: adding words, voices, and languages, mapped out — NERDDISCO · 2026-11-20
- Kale TTS roadmap: mlx-audio, llama.cpp and audio.cpp integrations already running — NERDDISCO · 2026-11-20
- [source] kale-tts-1 released open-weight in base/tiny/nano sizes, dataset included — NERDDISCO · 2026-11-20
- CPU-runnable open TTS models are piling up, starting with open-weights kale-tts-1 — Kyrannio · 2026-11-21
1 near-duplicate retellings: NERDDISCO