Is Underdog Woof state of the art?Pending benchmarks.Announced Nov 6, 2026; no independent results yet.
Underdog's tiny English text-to-speech family for local inference: a one-voice 30.7M model (31.7 MB) in the browser/WebGPU demo and a two-voice 38.2M model (40.6 MB) publicly downloadable from ConwayResearch/Underdog-Woof. The two-voice runtime uses numpy and outputs 24 kHz mono audio. The site and release post advertise CC BY-NC 4.0, but the repository's actual LICENSE is Underdog Non-Commercial License 1.0; commercial use requires a separate license.
The details so far
- Lab
- Underdog
- Announced
- Nov 6, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 30.7M one voice / 38.2M two voices
- Role
- English text-to-speech for local devices and browser inference
- Variants
- One voice: 30.7M parameters / 31.7 MB. Two voices: 38.2M parameters / 40.6 MB
- Access
- Ungated public weights at ConwayResearch/Underdog-Woof; public browser/WebGPU demo at underdog.ai/escape
- Runtime
- Two-voice numpy runtime; text normalization for numbers, dates, prices and codes; 24 kHz mono output
- License
- Repository LICENSE: Underdog Non-Commercial License 1.0 (HF tag other); commercial use needs a separate license. Site/X instead advertise CC BY-NC 4.0 — licensing discrepancy, not Apache-2.0
- Related
- Separate from Underdog Bark 3.0 and Suno Bark
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: WER on 500 hard sentences is 2.06% for one-voice Woof and 1.42% for the better of the two released voices; lab's comparison gives Fish Audio Drama 3 1.89%. These are lab measurements, not independent board results.
Lab claim: one-voice int8 model generates 13.2 seconds of speech in 0.73 seconds on one Apple M5 Max CPU thread at 16 steps.
Sources
LABUnderdog: public Woof releasex.com/UnderdogAI/status/2118921155713372212 · Nov 6, 2026LABUnderdog: browser WebGPU demox.com/UnderdogAI/status/2118921157500141961 · Nov 6, 2026LABUnderdog Woof model pageunderdog.ai/woofLABUnderdog Escape demounderdog.ai/escapeREPOConwayResearch/Underdog-Woof model cardhuggingface.co/ConwayResearch/Underdog-WoofREPOWoof actual license termshuggingface.co/ConwayResearch/Underdog-Woof/blob/main/LICENSETracking continues until an independent board lists it.