Underdog Woof first look: a 32MB open TTS model on CPU

By LLMTest Lab · Nov 7, 2026 · 4 min read voiceopen-weightslocal-llm
On this page

On this page

  1. What is new
  2. The vendor-reported numbers
  3. License is the real limit
  4. Who should care

Not tested by us yet. Woof ships as downloadable weights with no hosted API we can call, and we have not run it ourselves. Everything below comes from the model card on Hugging Face and TechCrunch's launch coverage. We'll update this post when we have compared its output against Kokoro on our own sentences.

Underdog Woof is a tiny English text-to-speech model from the startup Underdog that runs on a plain CPU with only numpy installed. The smallest build is a 31.7 MB download, and the vendor reports a 1.42% to 2.06% word error rate on its own hard-sentence test, ahead of Kokoro 82M at 5.17%. The catch is the license: it is free for personal and research use, and anything commercial needs a paid agreement.

Underdog Woof
Maker Underdog (ConwayResearch on Hugging Face)
Announced 2026-11-06
Parameters 38.24M total (22.95M speech model, 15.3M audio decoder)
Download 40.6 MB with two voices, 31.7 MB with one voice
Voices Juno (default) and Mira
Language English only
Output 24 kHz mono
Speed (Apple M5 Max, one CPU thread) Juno 13.8x real time, Mira 7.2x, one-voice build 18x
Peak memory 402 MB (Juno), 601 MB (Mira), 852 MB (one-voice build)
License Underdog Non-Commercial License 1.0

What is new

Most small TTS models trade a lot of quality for size. Woof's pitch is that it does not. The decoder is adapted from Kyutai's Mimi codec, and the model was trained on LibriTTS-R and Emilia-YODAS, both CC-BY 4.0 per the card. TechCrunch notes that the company expects the model to power local assistants or personalized podcasts rather than replace larger cloud voices, and that it also runs in the browser at underdog.ai/woof.

A 31M-class model that runs offline matters for one reason: no per-character fee, no network round trip, and text never leaves the device. If you are choosing between a cloud voice and a local one, our guide to privacy-first LLMs for apps covers the same trade-off for text models.

The vendor-reported numbers

All of the following are from the model card, scored on Underdog's own pipeline over 500 English sentences from the Gradium TTS Hard Cases set. The card says transcripts came from qwen3.8-omni-flash and errors were counted by gemma-4-31b-it.

Model Word error rate Download
Woof, two-voice (Mira) 1.42% 40.6 MB
Woof, two-voice (Juno) 1.53% 40.6 MB
Fish Audio Drama 3 1.89% 41.5 GB
Woof, one-voice 2.06% 31.7 MB
Phonon 31m 2.16% 55.1 MB
Kokoro 82M 5.17% 327 MB
Kitten Nano 0.8 9.81% 27.6 MB
Qwen3-TTS 1.7B 11.99% 4.54 GB

Read this with care. Word error rate here measures whether a speech recognizer can transcribe the output, so it tells you about intelligibility and pronunciation of hard sentences, not about how natural the voice sounds. The benchmark is run by the maker, the judge pipeline is theirs, and the card admits that one comparison (Liquid AI's LFM3-Audio-2B at 20.1%) used a different transcription method and is probably a little high. Treat the ranking as a claim to reproduce, not a result.

License is the real limit

Woof uses the Underdog Non-Commercial License 1.0, per the card. Free uses include personal projects, research, teaching and non-profits. You must credit "Underdog Woof" or "underdog.ai" in published work, and show "Voice by Underdog Woof" in any app that uses it. Paid products, ads, monetized video and company work need a license from Underdog. You also may not use Woof or its audio to train, fine-tune or distill another speech model. TechCrunch's article does not state the license at all, so check the card before shipping anything.

Who should care

  • Hobbyists and researchers who want a speech model that installs in one command and runs on a laptop. The 1.5% to 2% error rates, if they hold up, are good for the size.
  • Product teams should treat this as a licensing conversation first. If you need an open commercial voice, compare it with the options in our on-prem open-source model guide before committing.
  • Anyone needing more than English or more than two voices. Woof does not offer either.

Mira is slower and heavier than Juno, and long text is split and re-generated when the output looks wrong, so latency can vary from the figures above. Underdog's other releases include a 167 MB speech-to-text model called Bark, per TechCrunch. For another small open model that launched the same week, see our Alpie Seed explained post.

Related articles

Alpie Seed explained: a 9B Apache coding model for laptops
169Pi's Alpie Seed is a 9B Apache 2.0 model built on Qwen3.5-9B that runs offline. Specs, vendor benchmark claims, hardware needs and who should care.
Best LLMs for privacy-first apps in 2026: HIPAA, ZDR, on-prem
Which LLMs meet HIPAA, SOC 2, and data-residency requirements in 2026? Four paths compared: ZDR API, SOC 2 cloud, EU data residency, and full self-hosting.