underdog woof tts: A 40MB Voice Model That Runs on One CPU Thread

The underdog woof tts release is the kind of AI drop that makes cloud voice APIs look bloated. Underdog's Woof is an English text-to-speech model that fits in a 31.7 MB download for the one-voice browser version, or 40.6 MB with two voices, and runs on a laptop CPU with nothing but numpy. Underdog claims it beats models more than a thousand times its size on word error rate. That claim comes from Underdog's own testing, so read it with a raised eyebrow.

What actually happened

Underdog published the model card under the ConwayResearch organization on Hugging Face, and aggregators logged the release around November 6 and 7. The card describes a small English TTS model built to read phone numbers, dates, prices and codes "the way people say them," with a built-in rule-based text normalizer that needs only the Python standard library.

The specs are refreshingly concrete. The two-voice model totals about 38.24M parameters: 22.95M in the speech model and 15.3M in an audio decoder adapted from Kyutai's Mimi codec. It outputs 24 kHz mono audio. On one Apple M5 Max CPU thread, Underdog says Juno runs about 13.8x faster than real time with roughly 402 MB of peak memory, and the one-voice browser model hits about 18x real time.

AI Tamer, which covered the earlier browser demo, flagged two things worth knowing: the speed and offline claims come from the vendor, and the license is non-commercial, not Apache.

How the benchmark works, and why to be careful

Underdog ran every model in its comparison itself, scoring 500 English sentences from Gradium's hard-case set. Transcripts came from one model and word errors were counted by another, per the card. That is a reasonable automated setup, but it is still a vendor-run benchmark on a test set chosen by the vendor.

The numbers it reports are striking anyway. Woof's two-voice model posts 1.42% WER, the one-voice model 2.06%. Kokoro 82M, a popular on-device option, scored 5.17% on the same run. ElevenLabs Flash v2.5 came in at 6.63%. The only smaller download in the table, Kitten Nano 0.8 at 27.6 MB, scored 9.81%.

A tiny model winning a word-error test does not mean it sounds the most human. WER measures whether words come out right, not whether the voice feels natural.

If you follow the voice race, our heygen voice tts and open tts leaderboard pieces cover how naturalness and accuracy trade off across bigger models.

Why the underdog woof tts size matters

Small models change where voice can live. A 40 MB file that runs on a CPU thread can ship inside a desktop app, a browser extension or a game without a GPU, a server or an API key. That means lower latency, offline use and no per-character billing.

It also matters for AI agents. Agents that call, read notifications aloud or narrate dashboards need cheap, reliable speech. A model that handles "415-555-0199" and "5:30 p.m." correctly out of the box removes a lot of glue code.

The license fine print

The Hugging Face card calls it the Underdog Non-Commercial License 1.0, while earlier demo coverage described CC BY-NC 4.0. Either way, the rules are similar: free for personal projects, research, teaching and non-profits with credit; a commercial license for paid products, ads or monetized videos; and no using Woof or its audio to train other speech models.

So the underdog woof tts model is not a drop-in replacement for a commercial voice API unless you buy a license.

What to do as a reader

Try the browser demo and listen to the sample clips on the Woof Hugging Face card before believing any leaderboard, including Underdog's. Test it on your own text, especially names and numbers from your domain.

If you plan to build on it, read the license section twice and budget for a commercial deal if money is involved. Nothing here is investment advice.