In this categoryLocal AI · 75
More guides
Local AIBeginner

How to Run ConwayResearch/Underdog-Woof Locally: A 38M-Parameter CPU Text-to-Speech Model

Underdog Woof is a 38.24 million parameter English text-to-speech model from Conway Research, released November 5, 2026 under a non-commercial license. The whole model is a single 40.6 MB file with two voices, runs on a laptop CPU with nothing but numpy, and scores a 1.42% word error rate that the model card says beats models over 1,000 times its download size.

6 minBeginner

Underdog Woof is a 38.24 million parameter English text-to-speech model from Conway Research, released November 5, 2026 under a non-commercial license. The whole model, both of its voices included, is a single 40.6 MB file that runs on a laptop CPU using nothing but numpy, and it scores a 1.42% word error rate on a 500-sentence benchmark, which the model card says beats Fish Audio Drama 3, a model over 1,000 times its download size.

Conway Research built Woof for the company's own product, underdog.ai, and published the weights alongside it. The pitch is squarely about efficiency rather than raw capability: two voices, English only, in a model small enough to ship inside a browser tab or a phone app, with no server, no API key and no GPU anywhere in the pipeline.

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 Underdog Woof model page shows 33 likes and 10 downloads, two days after release. The most-liked linked Space is the Open TTS Leaderboard, a community benchmark board with 75 likes, not a demo Conway Research built itself. Conway's own demo Space, hugging-apps/underdog-woof, has no likes yet, and the model card links no GitHub repository, so there are no stars to report.

Specs at a glance

FieldValue
PublisherConway Research (ConwayResearch)
PipelineText-to-speech
Parameters38.24M total (22.95M speech model, 15.3M audio decoder)
Weights40.6 MB, both voices included
Output24 kHz mono
VoicesJuno (default) and Mira, both English
Audio decoderAdapted from Kyutai's Mimi codec
Training dataLibriTTS-R and Emilia-YODAS (CC-BY 4.0)
LicenseUnderdog Non-Commercial License 1.0 (custom, non-commercial)
Release dateNovember 5, 2026
Non-commercial license, with a credit requirement
Woof ships under Conway Research's own Underdog Non-Commercial License 1.0, not a standard Creative Commons or OSI license. It is free for personal projects, research, teaching and non-profits, but anything that earns money, including an ad, a monetized video or a paid product, needs a commercial license from Conway Research. Free use also requires crediting "Underdog Woof" or "underdog.ai" in what you publish, and the license bars using Woof's output to train, fine-tune or distill another speech model.

How small is 40.6 MB, really

Most TTS models built for quality run from tens of millions of parameters up into the billions, and a lot of them lean on a GPU or a cloud API to get there. Woof's whole model, two voices and all, fits in 40.6 MB and runs on one CPU thread. Conway Research built a one-voice variant too, a 31.7 MB download that also runs inside a browser tab at underdog.ai/woof, trading a little accuracy for an even smaller footprint.

Underdog Woof is a small English text-to-speech model with two voices: Juno and Mira. It reads phone numbers, dates, prices and codes the way people say them, and it runs on a laptop CPU with nothing but numpy, from a single 40.6 MB file.

Underdog Woof model card README (Hugging Face)

Word error rate against 18 other TTS models

Conway Research ran its own 500-sentence benchmark, Gradium TTS Hard Cases, against 18 other text-to-speech systems, scoring each in its best voice and crediting its own transcription and scoring pipeline (qwen3.8-omni-flash for transcripts, gemma-4-31b-it for counting errors). Woof's two-voice model, run in Mira, comes out on top of the whole list.

ModelRunsDownloadWord error rate
Underdog Woof, two voices (Mira)your CPU40.6 MB1.42%
Fish Audio Drama 3cloud API41.5 GB1.89%
Underdog Woof, one voiceyour CPU or browser31.7 MB2.06%
Phonon 31mon device55.1 MB2.16%
Kokoro 82Mon device327 MB5.17%
Pocket TTSon device219 MB8.07%
Supertonic 3on device398 MB8.81%
Kitten Nano 0.8on device27.6 MB9.81%
Qwen3-TTS 1.7Bon device4.54 GB11.99%

Kitten Nano is the only model in the full comparison with a smaller download than Woof, and it gets 9.81% of words wrong against Woof's 1.42%. Several cloud-only models sit in between, including Deepgram Flux TTS at 2.39% and ElevenLabs Flash v2.5 at 6.63%, but none of them publish a download size since they only run as an API. Conway Research also ran Liquid AI's LFM3-Audio-2B through a separate pass of the same 500 sentences, using Whisper large-v3 for transcripts since the preferred transcription API wasn't available for that run, and it scored 20.1% word errors at a 1.21 GB download.

Running it

Woof's only dependency is numpy. There is no PyTorch, no ONNX Runtime and no GPU driver to install.

bash - install and run Underdog Woof
$pip install numpy
$hf download ConwayResearch/Underdog-Woof --local-dir underdog-woof
$cd underdog-woof
$
$python -m underdog_woof "Your order ships on March 3rd. Call 415-555-0199 with questions." -o hello.wav
$python -m underdog_woof "Your order ships on March 3rd. Call 415-555-0199 with questions." --voice mira -o mira.wav
$

From Python, the Woof class takes the same text and voice options, plus a seed for reproducible output.

say.py
from underdog_woof import Woof

tts = Woof(voice="mira")   # "juno" (default) or "mira"
audio = tts("Hey! Thanks for calling. How can I help you today?", seed=0)   # float32 mono, 24 kHz
tts.save("hello.wav", audio)
It normalizes numbers, dates and codes for you
Woof ships a built-in rule-based normalizer that rewrites numbers, dates, times, money, codes and symbols the way people actually say them, plus a lexicon for respelling hard names. Type text the way you'd write it, not the way you want it read aloud. Pass normalize=False to Woof() to skip this and handle the text yourself.

Long input gets split into sentences, and anything over 120 characters is split again at its commas, semicolons or dashes. Each piece is checked as it's generated: if it runs too long for its text or has a stretch of silence in the middle, Woof regenerates it, up to three more tries, and keeps the best result. That retry logic is what the benchmark numbers above were measured with, one seed per sentence, so a fixed seed won't always reproduce if a piece happened to get redrawn.

License and use

Woof ships under Conway Research's own Underdog Non-Commercial License 1.0, which is listed as license type "other" on Hugging Face rather than a standard license. Personal projects, research, teaching and non-profit use are free, provided you credit "Underdog Woof" or "underdog.ai" in what you publish. A commercial license is required for anything that earns money, and Conway Research also prohibits using Woof's output to train, fine-tune or distill another speech model. Using Woof through underdog.ai itself is always free, commercial or not.

FAQ

What is ConwayResearch/Underdog-Woof?

It is a 38.24 million parameter English text-to-speech model from Conway Research, released November 5, 2026. It has two voices, Juno and Mira, stored in a single 40.6 MB file, and it runs on an ordinary CPU using only numpy.

Can I use Underdog Woof commercially?

Not without a separate commercial license from Conway Research. The Underdog Non-Commercial License 1.0 covers personal, research, teaching and non-profit use for free, with a credit requirement, but excludes anything that earns money.

Do I need a GPU to run it?

No. Underdog Woof runs on a single CPU thread with only numpy installed. On an Apple M5 Max, Conway Research measured 13.8 times real-time speed for the Juno voice and 7.2 times real-time for Mira.

How accurate is it compared to other TTS models?

Conway Research's own 500-sentence benchmark puts Woof's Mira voice at a 1.42% word error rate, ahead of every other model it tested, including the 41.5 GB Fish Audio Drama 3 at 1.89% and the much larger cloud APIs from Deepgram and ElevenLabs.

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
#underdog woof tts#conway research woof#cpu text-to-speech model#run underdog woof locally#small tts model comparison#numpy text to speech