Underdog Woof: Lightweight NumPy-Only English TTS Model

Sample outputs are available on the model card.
At a Glance
| Item | Value |
|---|---|
| Repository | ConwayResearch/Underdog-Woof |
| Family guide | Underdog-Woof guide (1 articles) |
| Published | 2026-11-06 |
| License | underdog-non-commercial-1.0 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
ConwayResearch has released “Underdog Woof," a lightweight English text-to-speech (TTS) model. By inputting English text, it outputs natural-sounding mono audio waveforms at a 24 kHz sampling rate.
This model features two speakers, Juno and Mira, and is designed to run entirely on CPUs such as laptops using only NumPy, requiring no GPUs, external servers, or API keys. It is equipped with the ability to properly convert phone numbers, dates, prices, and code notations into formats spoken by humans before generating speech.
Specifications
The main specifications of Underdog Woof as listed in the release model card are as follows:
- Parameter count: 22.95M for the speech model and 15.3M for the audio decoder, totaling 38.24M.
- Architecture: The audio decoder adopts a structure adapted from Kyutai’s “Mimi" codec.
- Output specifications: Outputs 24 kHz mono audio (float32 format).
- Supported languages: English is supported.
- Included speakers: Two speakers, Juno (default) and Mira, are included in a single weights file.
- Text processing features: Built-in rule-based normalization processing that operates using only standard libraries, handling the conversion of numbers, dates, times, currencies, and symbols, as well as dictionary conversion for hard-to-read personal names.
- Long-text processing and automatic regeneration: Long texts are split into sentence units, and sentences exceeding 120 characters are further subdivided by commas, semicolons, dashes, etc. If the generated audio is excessively long or contains silent intervals of 1 second or more, it regenerates up to 3 times to retain the optimal take.
- License conditions: The “underdog-non-commercial-1.0" license is applied, limiting usage to non-commercial purposes.
Performance and Quality
Using the evaluation dataset “Gradium TTS Hard Cases" (500 sentences) consisting of difficult English texts, the publishers evaluated existing open-weight models and cloud API models in the same environment and published a comparison of word error rates (WER: word-level error rate in transcriptions, lower is better). Transcriptions were performed using qwen3.8-omni-flash, and error counting was performed using gemma-4-31b-it.
The comparison table with other models published in the model card is as follows:
→ Scroll horizontally to see all columns
| Model | Runs | Download | Word error rate | Woof’s errors are lower by |
|---|---|---|---|---|
| Underdog Woof, two voices (Mira) | your CPU | 40.6 MB | 1.42% | |
| Fish Audio Drama 3 | cloud API | 41.5 GB open weights | 1.89% | 25% |
| Underdog Woof, one voice | your CPU or browser | 31.7 MB | 2.06% | |
| Phonon 31m | on device | 55.1 MB | 2.16% | 34% |
| Deepgram Flux TTS | cloud API | not public | 2.39% | 41% |
| Deepgram Aura-2 | cloud API | not public | 2.43% | 42% |
| Microsoft MAI-Voice-2.1 Flash | cloud API | not public | 2.51% | 43% |
| Fish Audio S2.1 Pro | cloud API | not public | 2.53% | 44% |
| Microsoft MAI-Voice-2.1 | cloud API | not public | 2.64% | 46% |
| Qwen Audio 3.0 TTS Plus | cloud API | not public | 2.74% | 48% |
| Mistral Voxtral 3 TTS | cloud API | not public | 2.91% | 51% |
| MiniMax Speech 2.8 Turbo | cloud API | not public | 4.63% | 69% |
| MiniMax Speech 2.8 HD | cloud API | not public | 4.88% | 71% |
| Kokoro 82M | on device | 327 MB | 5.17% | 73% |
| ByteDance Seed Audio 1.0 | cloud API | not public | 6.50% | 78% |
| ElevenLabs Flash v2.5 | cloud API | not public | 6.63% | 79% |
| Pocket TTS | on device | 219 MB | 8.07% | 82% |
| Supertonic 3 | on device | 398 MB | 8.81% | 84% |
| Kitten Nano 0.8 | on device | 27.6 MB | 9.81% | 86% |
| Qwen3-TTS 1.7B | on device | 4.54 GB | 11.99% | 88% |
Additionally, detailed measurement results for each of the two speakers included in the repository are as follows (speeds are medians measured over 3 rounds of everyday 200 sentences on a single Apple M5 Max CPU thread).
| Juno | Mira | |
|---|---|---|
| Word error rate, Gradium TTS Hard Cases (English) | 1.53% | 1.42% |
| Word errors | 67 in 4,371 | 62 in 4,379 |
| Speed on one Apple M5 Max CPU thread | 13.8x real time | 7.2x real time |
| Peak resident memory | 402 MB | 601 MB |
According to the published figures, Underdog Woof (Mira) has a word error rate of 1.42%, recording an accuracy that surpasses large-scale models and cloud API models from various companies, including Fish Audio Drama 3 at 1.89%. Among the comparison targets, Kitten Nano 0.8 is the only model with a smaller file size than Woof, but its error rate is 9.81%, and it is reported that no model smaller and more accurate than Woof was identified in the publisher’s tests.
On the other hand, there is a trade-off between accuracy and generation speed among the included speakers. While Mira achieves the lowest error rate at 1.42%, its speed on a single CPU thread remains at 7.2x real-time. In contrast, Juno has a slightly higher error rate of 1.53%, but enables faster speech generation at 13.8x real-time. Note that Liquid AI LFM3-Audio-2B, evaluated using Whisper large-v3, is reported to have had an error rate of 20.1% on the same sentence set.
Strengths and Use Cases
Underdog Woof specializes in synthesizing high-quality and high-accuracy English speech relying solely on general CPU environments such as laptops, without depending on external server communications or GPU resources. Its dedicated runtime design, which narrows down dependent libraries to just NumPy, allows for easy integration into edge devices, embedded systems, and local applications.
A major strength of this model is its advanced read-aloud processing tailored to formats humans speak daily. Mechanical strings such as phone numbers, order codes, dates, times, and currency notations are automatically converted into natural English phrasing by the built-in normalization function operating solely on standard libraries. Furthermore, because proper nouns that are difficult to pronounce are appropriately corrected by referencing a dictionary, it is extremely well-suited for conversational assistant use cases such as customer support automated response systems, order confirmation notification voice output, and smart speakers.
In addition, the two included speaking styles can be used selectively depending on the use case and hardware resources. The default Juno operates extremely fast even on a single CPU thread, making it suitable for real-time conversational responses and interactive voice interfaces where latency is prioritized. Meanwhile, Mira is suitable for producing important guidance announcements, audiobook narration, and long-form reading content where a lower word error rate is required.
Tolerance for long texts is also taken into account; even when long sentences exceeding 120 characters are inputted, they are automatically split and processed at comma and punctuation positions. Furthermore, if the generated audio waveform contains abnormal silent intervals or is excessively long in duration, a mechanism is provided to automatically regenerate up to 3 times to select a good-quality take, ensuring stable output even in unmanned batch processing or automated generation pipelines.
How to Get It
The weights and execution code for Underdog Woof are available from the Hugging Face repository. No prior application or license agreement procedures are required to use this model, and anyone can download it directly.
The only additional library required for operation is NumPy, allowing it to be used without installing PyTorch or heavy deep learning frameworks. The procedure to obtain the model from the official repository and synthesize speech from the command line (CLI) is as follows:
pip install numpy
hf download ConwayResearch/Underdog-Woof --local-dir underdog-woof
cd underdog-woof
python -m underdog_woof "Your order ships on March 3rd. Call 415-555-0199 with questions." -o hello.wav
python -m underdog_woof "Your order ships on March 3rd. Call 415-555-0199 with questions." --voice mira -o mira.wav
It is also possible to call it directly as a library from a Python script. By writing just a few lines of code as shown below, you can generate 24 kHz mono audio (float32 format) and save it as a WAV file.
from underdog_woof import Woof
tts = Woof(voice="mira") # "juno" (default) or "mira"
audio = tts("Hey! Thanks for calling. How can I help you today?", seed=0)
tts.save("hello.wav", audio)
If you want to perform text normalization yourself, you can disable the built-in normalization by specifying normalize=False when instantiating. Also, if you want to consistently reproduce audio with the same intonation and timing for the same text, pass seed as an argument to fix the seed value.
Can You Run It Locally?
License — underdog-non-commercial-1.0: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
We have not confirmed that this site meets the commercial-use terms of this model’s license (underdog-non-commercial-1.0). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-17 | ConwayResearch | MLX | ConwayResearch/Underdog-Woof-4B-1.1 | MLX 4bit 2.6GB (fits in 4GB VRAM) |
| 2026-11-17 | ConwayResearch | MLX | ConwayResearch/Underdog-Woof-2B-1.1 | MLX 4bit 1.6GB (fits in 4GB VRAM) |
File sizes of each build:
- Available builds in ConwayResearch/Underdog-Woof-4B-1.1: MLX 4bit 2.2GB
- Available builds in ConwayResearch/Underdog-Woof-2B-1.1: MLX 4bit 1.3GB
In addition, 3 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Other Models for the Same Task
Recent audio models covered by Local Model Watch, newest first. Grouped by the task each publisher declares on Hugging Face (pipeline_tag); the smallest VRAM tier is this site’s estimate.
- TTS-V5 Chinese TTS Model: 4GB+ VRAM, File List (4GB)
- TTS-V4 Chinese Text-to-Speech Model: 8GB+ VRAM, File List (1.9B, 8GB)
- tiny-voca English Text-to-Speech Model: 4GB+ VRAM, File List (4GB)
- TTS-V3 Phone-Dialogue Specialized TTS Model: 8GB+ VRAM, File List (1.9B)
What to Read Next
- Explore the same model family → Underdog-Woof family overview (1 articles, 2 converted builds)
- Formats this model is available in → MLX format guide and models
- How to read WER → Benchmark glossary
Sources
- ConwayResearch/Underdog-Woof (Hugging Face Model Card): https://huggingface.co/ConwayResearch/Underdog-Woof
- Underdog Woof Web Demo: https://underdog.ai/woof
- Gradium TTS Hard Cases Dataset: https://huggingface.co/datasets/gradium/tts-hard-sentences-202610
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: ConwayResearch/Underdog-Woof-4B-1.1, ConwayResearch/Underdog-Woof-2B-1.1

