In this categoryLocal AI · 75
More guides
Local AIBeginner

How to Run milliwattlabs/leokani-tts-7M Locally: A 7M-Parameter CPU Text-to-Speech Model

leokani-tts-7M is a 7 million parameter English text-to-speech model from Milliwatt Labs, released October 30, 2026 under a non-commercial license. It runs on an ordinary CPU with no GPU, ships as PyTorch and ONNX files as small as 4.9 MB, and already has Mac, Android, Linux, Home Assistant and browser builds.

6 minBeginner

leokani-tts-7M is a 7 million parameter English text-to-speech model from Milliwatt Labs, released October 30, 2026 under the cc-by-nc-4.0 license. It speaks in one female American English voice, runs on any CPU with no GPU required, and the smallest quantized build is under 5 MB. In two weeks it picked up Mac, Android, Linux, Home Assistant, Arduino and browser integrations.

Milliwatt Labs built it as a small, efficient alternative to running a full-size TTS model just to get a voice out of a smart speaker, a Raspberry Pi, or a home automation box. The name is Hawaiian: leo means voice, kani means sound. Under the hood, the single voice was trained on synthetic speech generated by Kokoro-82M, so it is effectively a distilled, much smaller descendant of that model rather than a fully independent voice.

Written by Priya Raghunathan, local-AI hardware reviewer. I verify parameter counts, licenses, hardware requirements and every command against the primary source before recommending a setup path.

At time of writing, the leokani-tts-7M model page shows 12 likes and 406 downloads. The linked demo Space has 1 like, and Milliwatt Labs' home-assistant-addons repo on GitHub has 7 stars. Small numbers for a model two weeks old, but the entry point is unusually low: a single CPU-only voice model that already runs on five different platforms.

Specs at a glance

FieldValue
PublisherMilliwatt Labs (milliwattlabs)
PipelineText-to-speech
Parameters7.0M total
Weights28 MB (fp32), 14 MB (fp16), 7.8 MB (int8 ONNX), 4.9 MB (int4 ONNX)
Output24 kHz mono
VoiceOne female American English voice
Trained onSynthetic speech generated by Kokoro-82M
Licensecc-by-nc-4.0 (non-commercial)
Release dateOctober 30, 2026
Non-commercial license
leokani-tts-7M ships under cc-by-nc-4.0, Creative Commons Attribution-NonCommercial. That rules out shipping it in a paid product or any commercial service without a separate agreement from Milliwatt Labs. Personal projects, Home Assistant setups and research are fine; a commercial app or SaaS product is not.

Why a 7M-parameter model, not a bigger one

Most open TTS models aimed at quality run from tens of millions to hundreds of millions of parameters, and plenty lean on a GPU to hit real-time speed. leokani-tts-7M goes the other way: one voice, trained by distilling from Kokoro-82M rather than from raw audio, so the model only has to learn to reproduce a single speaker instead of covering a wide range of voices and languages. That tradeoff is exactly what makes it small enough to run on a Raspberry Pi or inside a phone's system text-to-speech service.

Leokani-tts is a fast English text-to-speech model: high-quality speech in a tiny model for any CPU, no GPU needed.

leokani-tts-7M model card README (Hugging Face)

Running it with PyTorch

The repo ships both PyTorch safetensors and ONNX files, so there are two ways to run it from Python. The PyTorch path is the simplest to try first.

bash - install and run leokani-tts (PyTorch)
$hf download milliwattlabs/leokani-tts-7M --local-dir leokani-tts
$cd leokani-tts
$pip install -r requirements.txt
$
$python leokani.py "Hello there. How was your day?" -o hello.wav
$

Two flags are worth knowing right away: --speed adjusts pacing, and --fp16 halves the weight size for a small quality tradeoff.

bash - speed control and fp16 weights
$python leokani.py "A little slower, please." --speed 0.85 -o slow.wav
$python leokani.py --fp16 "Same voice, half-size weights." -o fp16.wav
$
Text-to-phoneme conversion needs espeak-ng, or a bundled alternative
By default leokani.py converts text to phonemes with espeak-ng, a separate GPL-3.0 dependency not included in the repo. Install it with brew install espeak-ng on macOS or sudo apt install espeak-ng on Debian and Ubuntu, or point to an existing install with --espeak PATH. To skip that dependency entirely, pass --g2p leokani to use Milliwatt Labs' own bundled phonemizer, Leokani G2P, included in the repo's g2p folder.
bash - run without installing espeak-ng
$python leokani.py "Hello there. How was your day?" --g2p leokani -o hello.wav
$

Running it with ONNX Runtime

For a lighter dependency footprint, or to use the int8 and int4 quantized files, run the ONNX build instead of PyTorch.

bash - run leokani-tts with ONNX Runtime
$pip install onnxruntime
$python leokani_onnx.py "Hello there. How was your day?" -o hello.wav
$python leokani_onnx.py "No espeak-ng needed." --g2p leokani -o g2p.wav
$python leokani_onnx.py --int8 "Same voice, a smaller file." -o int8.wav
$python leokani_onnx.py --int4 "The smallest file, with a small quality cost." -o int4.wav
$

The int8 file is roughly half the size of fp16 and, according to Milliwatt Labs' own blind listening tests, rated the same as full precision. The int4 file is the smallest download at 4.9 MB, with the model card flagging a small quality cost for that size.

Where it already runs

Past the command line, Milliwatt Labs has shipped integrations faster than most new TTS releases. A Mac app and a command-line tool with an MCP server cover desktop and agent use. An Android app doubles as the phone's system text-to-speech voice. Linux and Raspberry Pi builds cover x86_64 and ARM64. For home automation, adding the home-assistant-addons repository in the Home Assistant app store installs Leokani as an add-on for both amd64 and arm64, alongside a Wyoming protocol server and an OpenAI-compatible speech API that both run in Docker. A JavaScript library covers browsers and Node.js, and the demo Space runs the model entirely client-side in the browser, so typed text never leaves the device.

The most recent addition, shipped November 3, 2026, is an Arduino library for the ESP32-S3 and ESP32-P4, putting the model on microcontroller-class hardware. The same date brought Leokani G2P 1.2.0, an update to the bundled phonemizer that handles numbers, abbreviations, letters and homograph words more accurately, and a smaller int4 ONNX file for a faster download.

FAQ

What is milliwattlabs/leokani-tts-7M?

It is a 7 million parameter English text-to-speech model from Milliwatt Labs, released October 30, 2026 under the cc-by-nc-4.0 license. It speaks in a single female American English voice and runs on CPU without a GPU.

Can I use leokani-tts-7M commercially?

No, not without a separate agreement with Milliwatt Labs. The cc-by-nc-4.0 license covers personal, educational and research use, but excludes commercial use.

Do I need a GPU to run it?

No. leokani-tts-7M is built to run on an ordinary CPU, including on a Raspberry Pi or inside a Home Assistant add-on. Weight files range from 28 MB down to 4.9 MB depending on precision.

Does it need espeak-ng installed?

Only if you use the default text-to-phoneme path. Pass --g2p leokani instead to use Milliwatt Labs' own bundled phonemizer and skip the espeak-ng dependency entirely.

Where to go from here

Local models run better with more VRAM. CompareRTX GPUs on Amazonbefore you upgrade.(affiliate link. We may earn a commission at no extra cost. Disclosure)

Watch related tutorials

Free weekly email

Weekly local AI drops

New models, what runs on your hardware, and the guides to set them up. One email a week, unsubscribe any time.

Tags
#leokani-tts 7m#milliwatt labs tts#cpu text-to-speech model#run leokani-tts locally#kokoro cpu tts alternative#home assistant text-to-speech