Underdog Woof: A small English speech synthesis model that runs on CPU only

An English text-to-speech model by Conway Research (Underdog), featuring two voices, Juno and Mira. The entire model is a single 40.6 MB file that depends only on numpy. It includes a built-in rule-based normalizer that reads phone numbers, dates, times, amounts, and code in a conversational manner; long texts are automatically segmented into sentences, and generation anomalies trigger up to three automatic retries. The author reports a word error rate of 1.42% (Mira) on the English Gradium TTS Hard Cases set, 13.8 times real-time speed (Juno) on a single CPU thread of an Apple M5 Max, with a total of approximately 3824 ten thousand parameters, and output at 24 kHz mono.

Original title: ConwayResearch/Underdog-Woof

An English text-to-speech model by Conway Research (Underdog), featuring two voices, Juno and Mira. The entire model is a single 40.6 MB file that depends only on numpy. It includes a built-in rule-based normalizer that reads phone numbers, dates, times, amounts, and code in a conversational manner; long texts are automatically segmented into sentences, and generation anomalies trigger up to three automatic retries. The author reports a word error rate of 1.42% (Mira) on the English Gradium TTS Hard Cases set, 13.8 times real-time speed (Juno) on a single CPU thread of an Apple M5 Max, with a total of approximately 3824 ten thousand parameters, and output at 24 kHz mono.

Why it matters No GPU or API Key required. It correctly reads phone numbers, dates, and amounts in a conversational manner, making it suitable for offline reading and local applications.

Score increase 125 · Likes 126 · Downloads 10

Reasons

  • No GPU, server, or API Key required, suitable for offline and embedded scenarios
  • The author provides a comparison table using the same metrics as several open-source and cloud TTS models, along with samples that can be listened to directly
  • A single-voice version is available for online trial in the browser

Usage

  • Run pip install numpy, download the repository, then execute python -m underdog_woof "text to read" -o hello.wav. Use --voice mira to switch voices
  • In Python, generate using Woof(voice="mira"). Passing a seed ensures reproducible results

Risks

  • Supports English only
  • Non-commercial license: Commercial use requires a separate license, published works must include attribution, and using it or its audio to train other speech models is prohibited
  • Comparative evaluations were run and scored by the author themselves
  • This entry has not been tested; refer to the model card for actual performance

Open

Hugging Face Trending

Explore the latest AI Radar updates