What is Underdog Woof?
Underdog Woof is a small English text-to-speech model from Conway Research, the a16z-backed company behind the Underdog on-device assistant, published on 5 November 2026. It has two voices, Juno and Mira, both stored in a single 40.6 MB int8 weights file, and makes 24 kHz mono audio. It has 38.24 million parameters: a 22.95-million-parameter speech model and an audio decoder adapted from Kyutai's Mimi codec.
It runs on a laptop CPU with nothing but numpy. A built-in rule-based normaliser rewrites phone numbers, dates, prices and codes the way people say them, long text is split into sentences, and each piece is redrawn up to three more times if it runs long or contains a silence. A one-voice 31.7 MB version runs in the browser at underdog.ai/woof.
The licence is Underdog's own Non-Commercial License 1.0. Personal, research, teaching and non-profit use is free with credit; any commercial use, including a company's internal work, needs a paid licence whose price is not published. This model is English only and does not clone voices; a separate Japanese Woof appeared on Hugging Face on 10 November.
How much does Underdog Woof cost?
| Plan | Price | What you get |
|---|---|---|
| Non-commercial | Free | Personal, research, teaching and non-profit use, with credit |
| Commercial licence | Not published | Any business or money-making use; 30-day internal evaluation allowed |
Free for non-commercial use with credit under the Underdog Non-Commercial License 1.0. Commercial licence required for any business use; price not published. Free inside Underdog's own apps.
Is Underdog Woof worth it?
Woof's claim is accuracy at a size that barely registers. On Gradium's TTS Hard Cases, 500 English sentences full of numbers, codes and awkward names, Underdog reports a 1.42% word error rate for the Mira voice, against 5.17% for Kokoro 82M, 9.81% for the only smaller model it tested, Kitten Nano 0.8, and 1.89% for Fish Audio Drama 3, a 41.5 GB download. That is a model 1,022 times smaller getting fewer words wrong.
Treat those figures as a claim. Underdog ran and scored every model itself: qwen3.8-omni-flash transcribed the audio and gemma-4-31b-it counted the errors. Word error rate measures whether the words come out right, not whether the voice sounds natural, and no independent listening test or arena ranking has been published.
The speed is the practical win. The Juno voice runs 13.8 times faster than real time on a single CPU thread of an Apple M5 Max, peaking at 402 MB of memory, and Mira at 7.2 times. No GPU, server or API key is involved, so it suits offline apps, kiosks and accessibility tools on modest hardware.
The catch: the licence. Any use by or for a company counts as commercial, including internal operations, and needs a written licence from Underdog at a price it does not publish. Every published clip must credit Underdog Woof, apps must show "Voice by Underdog Woof", and you may not use its audio to train another speech model. There are two English voices and no voice cloning.
Our take: for a hobby project, a research prototype or a non-profit reader app, Woof is the most accurate voice under 50 MB on paper. For a product, Kokoro's Apache 2.0 licence still wins the argument before quality does.
Where it earns its place
A developer of a free, offline e-reader app for a literacy charity
Read books and notices aloud on cheap Android tablets with no network
Ship the 40.6 MB two-voice file (the 31.7 MB one-voice model runs only in Underdog's browser demo); non-profit use is free if the app shows "Voice by Underdog Woof" with a link. Community ONNX and Rust ports exist but are unofficial.
A speech researcher benchmarking tiny TTS models
Compare intelligibility of sub-100 MB models on hard text
Run python -m underdog_woof over the 500 Gradium sentences and score with your own ASR, since Underdog's 1.42% used its own pipeline. Research use is free; training another speech model on Woof audio is not allowed.
A product manager at a call-centre software firm
Read back order numbers and phone numbers in an on-premises voice bot
The built-in normaliser handles numbers such as 415-555-0199 and times like 5:30 p.m. Company use is commercial, so budget for a licence from license@conway.tech after the 30-day evaluation.
Before you commit
- Setup
- pip install numpy, download the repository with hf download, then run python -m underdog_woof "text" -o out.wav. Pass a seed for repeatable audio.
- Integrations
- Python API (Woof(voice="mira")) returning float32 24 kHz audio; command line. Community projects add a Rust port with an OpenAI-style server, a Docker HTTP API and Android ONNX builds, none endorsed by Underdog.
- Data and privacy
- Runs locally, so text never leaves the machine. The licence bans commercial use without a written licence, bans training speech or audio models on its outputs, and bans voice impersonation without consent.
- Cost at scale
- Free to run for non-commercial work on any CPU, at about 402 MB of memory per Juno process. Commercial licence fees are not published; ask license@conway.tech.
- Limits
- English only, two fixed voices, 24 kHz mono. Sentences over 120 characters are split at punctuation; no published naturalness score; trained on LibriTTS-R and Emilia-YODAS.
What’s good
- 40.6 MB for two voices; runs with only numpy
- 1.42% word error rate on Gradium hard cases (self-reported)
- 13.8 times real time on one CPU thread
- Reads numbers, dates and codes as people say them
What’s not
- Non-commercial licence; business use needs a paid licence, price not published
- Every benchmark run and scored by Underdog
- English only (Japanese is a separate model), two voices, no cloning
- Credit required on every published clip and app
Best for
Hobbyists, researchers and non-profits who want a clear English voice that runs offline on any laptop CPU.
Underdog Woof alternatives
Underdog Woof: common questions
- Is Underdog Woof worth it?
- Woof is the most accurate text-to-speech you can fit in 40 MB on Underdog's own test, and the licence keeps it out of almost any business use without a call to sales. We rate it 3.6 out of 5.
- How much does Underdog Woof cost?
- Free for non-commercial use with credit under the Underdog Non-Commercial License 1.0. Commercial licence required for any business use; price not published. Free inside Underdog's own apps.
- What are the downsides of Underdog Woof?
- Non-commercial licence; business use needs a paid licence, price not published; every benchmark run and scored by Underdog; english only (Japanese is a separate model), two voices, no cloning; credit required on every published clip and app.
- Who is Underdog Woof best for?
- Hobbyists, researchers and non-profits who want a clear English voice that runs offline on any laptop CPU.
- How much does Underdog Woof cost for a team or in production?
- Free to run for non-commercial work on any CPU, at about 402 MB of memory per Juno process. Commercial licence fees are not published; ask license@conway.tech.
- Is Underdog Woof safe for company data?
- Runs locally, so text never leaves the machine. The licence bans commercial use without a written licence, bans training speech or audio models on its outputs, and bans voice impersonation without consent.
- What are the best Underdog Woof alternatives?
- The closest alternatives we have reviewed are Kokoro-82M, KittenTTS 2, NeuTTS (Air, Nano, 2E) and Qwen3-TTS.
Answers checked Nov 11, 2026.