What happened
Underdog's ConwayResearch/Underdog-Woof page presents the repository's two-voice model, with Juno and Mira stored in one 40.6 MB weights file. The card distinguishes it from a one-voice version weighing 31.7 MB, which it says runs in a browser. The two-voice model runs locally on a CPU with NumPy; the listed installation uses pip install numpy and a Hugging Face download. The model produces 24 kHz mono audio. Its speech model has 22.95 million parameters and its audio decoder 15.3 million, for 38.24 million total.
The card says Woof reads phone numbers, dates, prices and codes in spoken form using a built-in rule-based normalizer, with an option to turn normalization off. It splits long text into sentences and further divides sentences longer than 120 characters at commas, semicolons or dashes. It checks each generated piece and may draw it again up to three times if it runs too long or contains at least a second of silence. A seed can be supplied to reproduce audio for the same text. The audio decoder is adapted from Kyutai's Mimi codec; the listed training data includes LibriTTS-R and Emilia-YODAS, both identified as CC-BY 4.0.
Underdog reports 1.42% word error rate for Mira and 1.53% for Juno on Gradium TTS Hard Cases, an English set of 500 sentences. Its comparison table says the two-voice Woof run used 40.6 MB and the one-voice build 31.7 MB. It reports 13.8 times real-time speed for Juno and 7.2 times for Mira on one CPU thread of an Apple M5 Max; peak resident memory is listed as 402 MB for Juno and 601 MB for Mira. These speeds are medians of three rounds over 200 everyday sentences. The card says the two-voice speed measurement uses its NumPy runtime. It compares those results with models including Fish Audio Drama 3, listed at 1.89% word error rate and 41.5 GB, while describing that model's run as a cloud API.
The license is Underdog Non-Commercial License 1.0. Personal projects, research, teaching and non-profits may use the model with credit, including publishing generated audio. The card requires attribution in published material and apps. Commercial use—including paid products, ads, monetized videos and company work—requires a separate license from Underdog. It also prohibits using Woof or its audio to train, fine-tune or distill another speech or audio model. The card says use through Underdog's own apps is always free; the Hugging Face page does not list an Inference Provider deployment.
What the archive adds
New Runtime's earlier coverage on November 5 described the initial browser-oriented Woof release: a 30.7-million-parameter model for local browser use. This is a new fact on a known story: the Hugging Face page now documents a two-voice CPU build and its own benchmark figures.
Trend signal
This release fits the pattern that generative media is becoming application infrastructure: Woof is presented as a local component with repeatable input handling, seeded output and deployment-specific builds, rather than only a one-off voice-generation demo. The evidence is about offline TTS, however, not real-time conversational voice systems, so it does not establish the broader pattern by itself.
Why now
The repository page now separates a browser-sized one-voice version from a two-voice CPU version, with distinct memory, speed and accuracy figures. For teams evaluating local speech, that makes the trade-off between smaller browser delivery and a larger CPU runtime more legible than a single model-size claim.
Why it matters
Local speech generation can reduce dependence on a hosted inference API and may make offline or privacy-sensitive interfaces easier to prototype. Woof's small download and reported CPU performance are relevant to applications that need spoken confirmations, generated prompts or numeric readouts without a GPU server. Its number-and-date normalization is especially pertinent where naive text-to-speech often misreads structured strings.
The measured result still needs to be treated as a vendor benchmark, not a general guarantee: the test set is English, and the card reports the model author's own runs and scoring setup. More immediately, the license determines whether many teams can ship it at all. A free model for personal, research or nonprofit work is not a free commercial dependency; paid products and company use require a separate agreement. Builders should assess legal fit alongside latency, memory and audio quality, and compare the exact version and voice they would deploy.
Limits and uncertainty
The performance and accuracy results are reported by Underdog; the card does not provide independent replication. The word-error comparison covers 500 English sentences, and the reported CPU speed uses an Apple M5 Max, so results may differ on other hardware and text. The page gives no commercial license price or terms beyond requiring a license. It also does not establish performance across languages: the model is described as English. No independent assessment of output quality, pronunciation edge cases or the license's application to particular products is included here.
Skill to build now
This week, take a representative set of your product's phone numbers, dates, prices, identifiers and ordinary prose, then run it through both Juno and Mira with fixed seeds. Record pronunciation failures, generation time and peak memory on your target CPU. That small test will show whether a local TTS component is technically plausible before you invest in integration—and whether the non-commercial license is already a blocker.
Reader takeaway
Woof makes CPU-based English speech generation look practical at a modest download size, with vendor-reported speed and numeric-reading results. The practical boundary is not just model quality: the published license is non-commercial, so a team building a paid product must secure separate permission before treating it as a shippable dependency.