█

AI/TLDR

Muhammed Nazeem (n4ze3m) · 2026-11-14 · notable

tiny-voca — a 4.5M-parameter open TTS model that runs 70x real time on CPU

In short

tiny-voca is an Apache-2.0 English text-to-speech model with about 4.5M parameters and two voices. It runs about 70x faster than real time on four CPU threads, and its ONNX file also runs in the browser.

Muhammed Nazeem (n4ze3m) · 2026-11-14 · 2 sources checked

Hugging Face model card preview for n4ze3m/tiny-voca

A text-to-speech model small enough to fit in 9 MB, with two English voices and no GPU needed.

What is it?

tiny-voca is a small English text-to-speech model from developer Muhammed Nazeem, with a female (F2) and a male (M2) voice and 24 kHz mono output. The weights take 18.2 MB in PyTorch, and GGUF quantizations go down to 8.7 MB. In the 16 November 2026 Open TTS Leaderboard snapshot, its M2 voice ranks 8th of 50 English models without voice cloning.

How does it work?

Instead of a neural vocoder, the model predicts pitch, voicing, energy and filter settings and feeds them to a classic source-filter synthesizer. A 4-block transformer reads phonemes from the dizione phonemizer, a duration predictor sets each phoneme's length, and two small frame-level transformers produce the synthesis parameters once every 256 samples.

Why does it matter?

Speech this cheap can run on any laptop, in a web page through onnxruntime-web, or inside small apps without a server. The trade-offs are stated plainly on the model card: no voice cloning, no streaming, only two fixed voices, and speech that can sound slightly synthetic.

Who is it for?

app developers who need offline or in-browser speech

Try it

pip install git+https://github.com/n4ze3m/dizione git+https://github.com/n4ze3m/tiny-voca

Sources · 2 outlets

Tags

  • tiny-voca
  • text-to-speech
  • tts
  • speech
  • on-device
  • cpu-inference
  • onnx
  • gguf
  • open-weights
  • apache-2-0

← All releases · Learn AI