tiny-voca is a small English text-to-speech model. It has two voices, 4.5M parameters, and runs on a CPU or in the browser.
- Small: 4,537,291 parameters, 18 MB, waveform generator included.
- Fast: about 70x real time on four CPU threads.
- Two voices: F2 (female) and M2 (male), American English, 24 kHz.
- Runs anywhere: PyTorch, ONNX Runtime, and ONNX Runtime Web.
- Open: code, weights, training recipe and dataset. Trained from scratch.
git clone https://github.com/n4ze3m/tiny-voca.git
cd tiny-voca
pip install -e ".[own]"
hf download n4ze3m/tiny-voca --local-dir weights
export DIZIONE_DATA=weights/phonemizerpython -m tiny_voca.infer \
--checkpoint weights/tiny-voca.pt \
--voice F2 \
--text "Hello! This is tiny voca." \
--output hello.wav| Option | Default | Meaning |
|---|---|---|
--text-file |
Read the text from a file instead of --text |
|
--voice |
F2 |
F2 or M2 |
--speed |
1.0 | Speaking rate |
--pitch-shift |
0 | Raise or lower the pitch |
--pitch-range |
1.0 | More or less pitch movement |
Long text is fine. It is read one sentence at a time and the audio is joined.
A text that ends without a mark gets a full stop, so the last word does not
break off. A sentence with nothing to read, such as one in another script, is
skipped. The output never peaks above 0.99, so it does not clip as a 16-bit
WAV. In Python, tiny_voca.infer.speak_text does the same.
python -m tiny_voca.export --checkpoint weights/tiny-voca.pt --output tiny-voca.onnxA ready file is in the weights repository as onnx/model.onnx. The
model card shows how to call it.
The demo runs the model in the browser. It
plays each sentence as soon as it is made, and keeps the whole clip in the
player. Its source is in web/.
cd web
bash fetch_assets.sh /path/to/dizione
npx wrangler devAn experiment in js/: the model and the phonemizer in plain scripts, with no
libraries. It reads one 8 MB file and runs in the browser, in Node and in Bun.
node js/say.mjs "Hello! This is tiny voca." --voice M2 --output hello.wavSee js/README.md.
export TINY_VOCA_FRONTEND=own
python -m tiny_voca.prepare --inputs data/train-*.parquet --split train --output features
python -m tiny_voca.annotate --features features --output annotations
python -m tiny_voca.train --train annotations/train.jsonl --validation annotations/validation.jsonl \
--output runs/NAME --steps 104000 --gan-start 8000 --gan-ratio 0.7 --teacher-final 0
python -m tiny_voca.evaluate --checkpoint runs/NAME/inference.pt --annotations annotations --output eval/NAMEThe released weights took 7.7 hours on one RTX 4090. More about the model and the data is in docs/model.md.
PYTHONPATH=src:tests python -m unittest discover -s tests- English only. Two fixed voices. No voice cloning.
- The F2 voice sounds younger than intended. A more mature voice is planned.
- Foreign names and odd spellings are often read wrong.
- No streaming yet.
Built by other people on top of tiny-voca.
Languages and voices:
- apptor-tiny-tts: German.
- tiny-voca-it: Italian.
- zerre-tts: Turkish.
- clarvia-tts-tiny: Turkish, ten voices, with its own phonemizer for numbers, dates and money.
- korean-tts: Korean, and Korean mixed with English.
- tiny-voca-ja-tsukuyomi-chan: Japanese, in Tsukuyomi-chan's voice, under that voice's own terms.
- slyszeglosy: Polish.
- mini-voz: Spanish, with a Castilian accent.
- tiny-voca-es: Spanish from Spain, with a browser demo.
- tiny-voca-fr: French, with a browser demo.
- tiny-voca-ptbr: Brazilian Portuguese, with its own phonemizer.
- tiny-voca-ptbr-kokoro: Brazilian Portuguese, with Kokoro's Brazilian voices as the teacher.
- tiny-voca Arabic: Modern Standard Arabic, trained on a CPU from the Arabic Speech Corpus.
- tiny-voca-ar: Modern Standard Arabic, an early version with a browser demo.
- tiny-voca-he: Hebrew, one male voice, with a browser demo.
- Bengali Voice: Bengali script and Romanized Bengali, with a browser demo.
- tiny-voca-hindi: Hindi, one voice with eight speaking styles, and a C++ engine.
- tiny-urdu: Roman Urdu, with a browser demo.
- Kurdish Behdînî Voice: Kurdish Behdînî, 21 voices, with a browser demo.
- alpa: Sanskrit, fine-tuned from tiny-voca on a CPU.
- fiato: English, with sighs, exhales and gasps.
Ports and apps:
- llama.cpp: runs tiny-voca with
llama-tts, from GGUF files. - voca: a rewrite in Rust.
- tiny-voca-esp32s3: runs on an ESP32-S3 microcontroller.
- tiny-voca-ane: runs on the Apple Neural Engine with Core ML, with a Swift package.
- tiny-voca-dashboard: a local dashboard.
- tinyvoca: a Mac app.
- tiny-voca-esp32s3: ESP32-S3 port.
Same idea, other sounds:
- tinygroove: a drum machine with no samples.
Write-ups:
- Three small TTS models: how tiny-voca works, measured next to Kitten TTS and Tontaube on one CPU.
- tiny-voca review: what it does well, where it breaks and how to run it.
Made something? Open an issue and I will add it.
The whole project cost $9.56 of GPU time on one rented RTX 4090. That is the dataset, the failed tries and the final training run.
It cost less than a pizza to make. If you like it, you can buy me one: https://ko-fi.com/n4ze3m
Apache-2.0 for code and weights. CC BY 4.0 for the dataset.
@software{nazeem2026tinyvoca,
author = {Nazeem, Muhammed},
title = {tiny-voca: a 4.5M-parameter two-voice English text-to-speech model},
year = {2026},
url = {https://huggingface.co/n4ze3m/tiny-voca}
}