The takeaway
A 31.7 MB on-device TTS model is now accurate enough for constrained input, which removes latency, per-character cost and privacy risk from voice agents. The blocker is the CC BY-NC license: prototype yes, commercial product no.
Why it matters for builders
Local TTS at a 40 MB footprint makes offline voice agents and on-device dictation practical: no network hop, no per-character bill, no audio leaving the machine. But Woof is CC BY-NC 4.0, so it is a prototyping and research asset, not a shippable dependency for a paid product. Benchmark it on your own inputs (codes, prices, names) before you treat 1.42% WER as a guarantee.
A 32MB voice model beats cloud TTS a thousand times its size
Most text-to-speech still happens in someone else's data center. Underdog, the on-device AI startup founded by Sigil Wen, argues it does not have to: its Woof model speaks English from a 31.7 MB file, inside a browser tab, with the Wi-Fi switched off.
TechCrunch reported the demo on November 6, built as a small retro-tech game: a room with a typewriter and a CRT monitor where whatever you type is spoken back to you. Everything runs locally.
What Underdog shipped
Woof comes in two sizes. The single-voice build is 31.7 MB with 30.7 million parameters and runs in the browser at underdog.ai/woof. The two-voice build, Juno and Mira, is 40.6 MB with 38.2 million parameters and is downloadable from Hugging Face.
The headline number is accuracy per byte. On 500 sentences packed with phone numbers, dates, prices and codes, the two-voice model scored a 1.42% word error rate; the one-voice build scored 2.06%. Fish Audio Drama 3, a 41.5 GB open-weight model, scored 1.89% on the same set. Underdog puts Woof at 1,022 times smaller than Drama 3 with 25% fewer word errors. Those benchmarks were run and scored by Underdog itself, so treat them as vendor numbers until someone reproduces them.
The engineering claim matters as much as the score. Woof's two-voice runtime is plain numpy, no GPU, no API key, 402 MB peak memory, and 13.8 times faster than real time on a single Apple M5 Max CPU thread. It lands weeks after Bark, Underdog's 167 MB speech-to-text model.
The catch for anyone building on it
Woof is open weights, but not open source in the Apache 2.0 sense. It ships under CC BY-NC 4.0 on the demo page and an "underdog-non-commercial-1.0" license on Hugging Face. You can read the weights and learn from them, but you cannot put them in a product you charge for without a separate deal. That single line decides who this release is for: a strong prototyping and research drop, not a drop-in replacement for a commercial TTS bill.
Why it matters
Voice is the most latency-sensitive part of an agent stack. A local 30 MB model removes the network hop, the per-character charge and the privacy question at once, with a footprint small enough for a desktop app. Underdog has shown the quality floor that makes that trade defensible for constrained input like codes, prices and dates.
The open question is no longer size. It is licensing and reproducibility: a vendor-run benchmark is a promise, not a guarantee, and the teams best placed to exploit a tiny voice model are the ones least able to use this one.
The Automation Brief
Read 5 AI stories instead of 50.
The essential moves in AI agents, models, automation and infrastructure — filtered for builders and operators, with the part that actually matters.
No noise. Unsubscribe anytime.
Editorial notes
Stefan Trbojevic
n8n Lab Editorial
6 November 2026
6 November 2026
Sources
AI disclosure: AI assisted with research and drafting. Factual claims are reviewed by an editor.
