Show HN: Three new Kitten TTS models – smallest less than 25MB

Original Article ↗ Hacker News Discussion ↗

Tiny new KittenTTS speech models, with the smallest under 25MB and open weights, are drawing interest for on-device and edge use cases ranging from mobile apps and browsers to Raspberry Pi and accessibility tools. Commenters praise the quality and expressiveness for the size but note issues with prosody, number pronunciation, cartoon‑like default voices, and heavy Python dependency chains that complicate installation. The creators signal an aggressive roadmap, including multilingual support, professional and custom voices, streaming output, mobile and browser SDKs, and future speech‑to‑text and voice cloning models.

Model capabilities & quality

  • New KittenTTS models range from ~15–80M params; smallest under 25MB.
  • Many users impressed by quality vs size, especially the 15M model outperforming the project’s previous 80M version.
  • Prosody and expressiveness are praised for a tiny model, but issues remain with numbers, abbreviations, and some technical terms.
  • Small models are said to struggle more with prosody; author claims next releases will further improve rhythm and intonation.

Voices, accents & customization

  • Current voices are described as somewhat “cartoon/anime/helium”; several people want more neutral, professional voices for business and audiobooks.
  • Requests for custom voice cloning, realistic accents, and more “serious” voices. A DIY custom-voice path and better pro voices are promised soon, plus a cloning model (15M–500M).
  • Interest in specific languages: Spanish, French, German, Japanese, Bengali, Norwegian, Irish/British/Welsh accents. Multilingual and Japanese support are said to be in progress.

Control, tags & expressivity

  • Strong demand for finer control: pitch/speed/volume, emotional tags ([sarcastically], [happily]), non-speech sounds ([gasp], [laughter], [clapping]), and intonation control.
  • Current version does not support expressive tags, but developer is considering a small, well-defined tag set and possibly streaming/chunked generation.

Platforms, deployment & APIs

  • High interest in on-device/mobile: iOS, Android (as system TTS for ebook readers, screen readers), Raspberry Pi, MCUs, browser/WebAssembly, C++/ONNX, and JS packages.
  • Mobile SDK, custom inference engine, streaming support, browser/edge SDK, and text–audio alignment output are all on the roadmap.
  • Edge/Next.js and Arduino-style chip deployment are seen as compelling due to the small model size.

Installation, dependencies & tooling

  • Many complaints about Python environment issues: huge Torch/CUDA dependency chain (multiple GB), version incompatibilities (Python 3.14, spaCy, misaki), and packaging bloat.
  • Community members share workarounds (CPU-only Torch, trimming unused imports, CLI wrappers) and ask for simpler, self-contained binaries or CLIs.
  • The maintainer acknowledges environment problems and promises to slim dependencies, add better env tooling (uv/conda), and fix packaging bugs.

Use cases & related wishes

  • Mentioned use cases: article/news/audiobook reading, screen readers, epub readers, automated business calls, home assistants, virtual pets, and offline accessibility.
  • Some ask about STT models; the team says they are working on small STT with better formatting (line breaks, quotes) rather than just low WER.