Skip to main content
MindStudio
Pricing
BlogAbout
My Workspace
Underdog Woof TTSrun TTS on CPUsmall text-to-speech model

Underdog Woof TTS: Run This 40MB Voice Model Locally on CPU

A hands-on look at Underdog Woof, a 40MB text-to-speech model that runs on CPU with no GPU, plus word-error-rate testing results.

Edited by Luis Chavez-Mattos, Director of Product RSS
Underdog Woof TTS: Run This 40MB Voice Model Locally on CPU

What is Underdog Woof TTS?

Underdog Woof is a small English text-to-speech model built to run on CPU-only hardware, with a total download size of around 40.6 MB. It ships with two voices, named Juno and Mira, and needs only NumPy to run, no GPU, no API key, and no heavyweight server infrastructure. It’s aimed at edge devices and lightweight local deployments rather than production voice pipelines that demand frontier-level accuracy.

TL;DR

  • Underdog Woof is a roughly 40.6 MB text-to-speech model that runs entirely on CPU using only NumPy as a dependency, with no GPU required.
  • It offers two voices, Juno and Mira, and the model card reports a 1.42% word error rate on a “TTS hard cases” benchmark.
  • Hands-on testing on an Ubuntu VM showed near-zero GPU usage and minimal CPU load, with inference times under five to six seconds per sentence.
  • Simple sentences come out clean, but complex inputs trip it up, including phone numbers, tracking codes, titles, abbreviations, and multi-part addresses.
  • On harder benchmark-style sentences involving big numbers, percentages, ordinals, and currency symbols, the model made two to three errors per sentence in testing.
  • The model is free for non-commercial use with attribution, but the commercial license is restrictive and not Apache 2.0, which limits its usefulness for production products.
  • Compared to similarly tiny models like Kitten Nano, Woof reportedly scores better on accuracy, though much larger models like Fish or Dia (referred to as “U-Drama 3” in testing) still outperform it in real-world listening tests.

4 hours from idea to a live business app.

Free live class · Wed, Dec 2
Remy Certified AI Engineer badge, Cohort 01

How do you install and run Underdog Woof on CPU?

The setup process is straightforward and doesn’t require specialized hardware. Based on the hands-on walkthrough, the general steps are:

  1. Create a virtual environment (an Ubuntu VM was used in testing, but any Linux environment with Python should work).
  2. Install the prerequisites, which for Underdog Woof is minimal since the model depends only on NumPy.
  3. Download the model files from its Hugging Face model card, which total around 40.6 MB.
  4. Serve the model using the provided command-line tool.
  5. Optionally, wrap it in a simple interface, such as a Gradio GUI, to interact with it through a browser instead of the command line.

Once running, the model draws essentially nothing from the GPU and only a small amount of CPU, which is the entire point of a model this size. This makes it realistic to run on modest VMs, small servers, or edge devices without dedicated AI hardware.

How accurate is it on real sentences?

The model card’s own benchmark reports a 1.42% word error rate on what it calls “TTS hard cases,” and claims the model can correctly read phone numbers, dates, prices, and quoted text the way a person naturally would. Hands-on testing largely confirmed the model performs well on simple sentences but starts to break down as complexity increases.

In one test, a sentence combining a date, time, phone number, invoice number, dollar amount, and tracking code was read in Juno’s voice with mostly correct output, though dashes in the phone number weren’t spoken and there was a noticeable pause or hiccup around the digit sequence. The same sentence in Mira’s voice took longer to generate and included more slips, particularly around the numbers.

A second test loaded a sentence with abbreviations, a professional title, a fraction, a unit of measurement, a date, and an email address. This produced several mistakes, described as roughly a coin-flip’s worth of errors, confirming that names, titles, and structured data like fractions and emails are a weak point.

A third test pushed further into financial and statistical language: big dollar figures, percentages, ordinals, decimals, and a Roman numeral. This produced two to three errors in a single sentence, reinforcing that the model handles conversational language better than data-dense or formal text.

Is Underdog Woof worth using?

For non-commercial, lightweight use cases, it’s a reasonable option if small size and zero GPU dependency matter more than top-tier accuracy. The model card’s benchmark chart plots model size against word-error rate, with the ideal zone being small and accurate. According to that chart, Woof claims to sit in that ideal corner, meaning models just as accurate (like Fish or Dia) are tens of gigabytes larger, while models closer to Woof in size (like Kitten Nano) reportedly make more mistakes.

Everyone else built a construction worker.
We built the contractor.

🦺
CODING AGENT
Types the code you tell it to.
One file at a time.
🧠
CONTRACTOR · REMY
Runs the entire build.
UI, API, database, deploy.

Hands-on testing partly supports this positioning but adds nuance. Fish and Dia, despite being far larger, produced noticeably better output quality in direct listening comparisons. Kitten Nano, meanwhile, performed similarly to Woof rather than clearly worse, suggesting the benchmark chart reflects best-case results on the model’s own test set rather than a universal ranking. As always with vendor-published benchmarks, independent testing is worth doing before relying on the numbers.

For commercial products, the picture changes. The license allows free non-commercial use with credit, but the commercial terms are restrictive and not based on Apache 2.0, which is a real blocker for anyone hoping to ship this inside a paid product without separate licensing arrangements.

What are the practical limitations?

The biggest limitation is handling structured or data-heavy text. Phone numbers, tracking codes, fractions, units, Roman numerals, percentages, and email addresses all introduce errors at a noticeably higher rate than plain conversational sentences. Pauses and small hiccups appear around digit-heavy sections, and voice choice matters too: Juno performed more reliably than Mira across the same test sentences in hands-on testing.

Generation speed is solid for the size, with most single-sentence outputs completing in under five to six seconds on CPU. That’s fast enough for many local, low-stakes applications like prototyping, accessibility tools, or simple voice notifications, but not necessarily fast or accurate enough for anything reading financial reports, legal documents, or contact information aloud without human review.

Frequently Asked Questions

What is the size of the Underdog Woof TTS model?

The full model download is approximately 40.6 MB, making it small enough to run on CPU-only systems and edge devices without specialized hardware.

Does Underdog Woof need a GPU to run?

No. It’s designed to run on CPU using only NumPy as a dependency. Testing confirmed essentially zero GPU usage and minimal CPU load during inference.

How many voices does Underdog Woof offer?

Two voices are included: Juno and Mira. In hands-on testing, Juno produced more consistent results on complex sentences than Mira.

Can I use Underdog Woof commercially?

It’s free for non-commercial use with attribution, but the commercial license is more restrictive and is not Apache 2.0, which limits its suitability for commercial products without further review of the license terms.

How accurate is Underdog Woof on difficult text?

The model card reports a 1.42% word error rate on its own hard-case benchmark. Independent testing found it handles simple sentences well but makes multiple errors on sentences packed with numbers, abbreviations, titles, or financial figures.

Editorial standards

Presented by MindStudio

No spam. Unsubscribe anytime.