Underdog Woof, a 40.6 MB two-voice TTS model, claims 1.42% word error rate on a CPU

Underdog Woof, an English text-to-speech model published on Hugging Face under ConwayResearch, fits both of its voices in a 40.6 MB file and runs on a laptop CPU with only numpy installed. Underdog reports a 1.42% word error rate for the Mira voice on Gradium TTS Hard Cases (English, 500 sentences).

Key specifications

• Voices: Juno (default) and Mira, both stored in one weights file
• Parameters: 22.95M speech model plus 15.3M audio decoder, 38.24M total
• Download: 40.6 MB
• Output: 24 kHz mono audio
• Peak memory on one CPU thread: 402 MB (Juno), 601 MB (Mira)
• Speed on one Apple M5 Max CPU thread: 13.8x real time (Juno), 7.2x real time (Mira)
• Pricing: not applicable for local use; commercial licensing terms not disclosed

The audio decoder is adapted from Kyutai's Mimi codec. Training data includes LibriTTS-R and Emilia-YODAS, both CC-BY 4.0.

Benchmark results (run by Underdog)

Underdog says it ran and scored every model itself, on the same 500 sentences, using each model's best-scoring voice. Transcripts came from qwen3.8-omni-flash, and gemma-4-31b-it counted the word errors. Lower is better.

Model	Runs	Download	Word error rate
Underdog Woof, two voices (Mira)	CPU	40.6 MB	1.42%
Fish Audio Drama 3	not stated as cloud or local in this row	41.5 GB	1.89%
Underdog Woof, one voice	CPU or browser	31.7 MB	2.06%
Phonon 31m	on device	55.1 MB	2.16%
Kokoro 82M	on device	327 MB	5.17%
Kitten Nano 0.8	on device	27.6 MB	9.81%
Qwen3-TTS 1.7B	on device	4.54 GB	11.99%

Underdog also lists cloud APIs from Deepgram, Microsoft, Fish Audio, Qwen, Mistral, MiniMax, ByteDance and ElevenLabs, with word error rates from 2.39% to 6.63% in its table. Cloud-only models do not publish a download size. Kitten Nano 0.8 is the only smaller download Underdog tested, and Underdog notes that nothing it tested is both smaller and more accurate than Woof.

Liquid AI's LFM3-Audio-2B (1.21 GB, 4-bit GGUF) scored 20.1% (747 of 3,718 words) in a separate run of the same sentences. That run used Whisper large-v3 transcripts because qwen3.8-omni-flash could not be used for it. Underdog says the figure is likely a little higher on its main setup, based on a Kokoro check that scored 4.29% with Whisper against 5.17% above.

One voice or two

A one-voice Woof is also available and powers the browser demo at underdog.ai/woof. It is 31.7 MB, uses 852 MB peak memory on one CPU thread, scores 2.06% word error rate and runs at 18x real time. Underdog says this speed is plain PyTorch with int8 weights at the default 16 steps. The two-voice speed uses the numpy runtime in this repo.

Text handling

A built-in rule-based normalizer rewrites numbers, dates, times, money, codes and symbols into spoken form, and respells hard names from a lexicon. It needs only the Python standard library and can be disabled with normalize=False. Long text is split into sentences, and sentences over 120 characters are split again at commas, semicolons or dashes. If a piece runs far too long or contains a second or more of silence, Woof redraws it up to three more times and keeps the best take. A seed argument makes output repeatable.

License

Woof is released under the Underdog Non-Commercial License 1.0. It is free for personal projects, research, teaching and non-profits, with credit to Underdog. Commercial use needs a license from Underdog (license@conway.tech). The license prohibits using Woof or its audio to train, fine-tune or distill another speech or audio model. No Inference Provider currently deploys the model.

What this means

The reported results are Underdog's own measurements, so they depend on its sentence set, transcription model and scoring setup. On those terms, Woof pairs a sub-50 MB download with a lower word error rate than every other model in the comparison, including the 41.5 GB Fish Audio Drama 3. The numpy-only runtime and CPU speeds of 13.8x (Juno) and 7.2x (Mira) real time make local, offline speech synthesis possible without a GPU or API key. The license limits the audience: commercial products need a separate agreement, and the ban on training other speech models rules out distillation uses.