Underdog-Woof: CPU-only TTS that is 13.8x faster with a 40.6MB size

A lightweight English text-to-speech model that runs on a single-threaded laptop CPU without a GPU, server, or API key. It supports two voices, Juno and Mira, and is distributed as a single file of 40.6MB. It runs with only numpy and generates 24kHz mono audio.

1 / 2

It includes a rule-based normalizer to read phone numbers, dates, prices, and code naturally, as a human would. Sentences exceeding 120 characters are automatically split at commas or semicolons, and the quality of the generated audio is verified, with defective segments retried up to 3 times. Providing the same seed always outputs the same audio.

It demonstrates overwhelming efficiency compared to competing models. It is 1,022 times smaller than Fish Audio Drama 3 (41.5GB) and has a 25% lower word error rate on the same test set. It processes 13.8 times faster than real-time on an Apple M5 Max CPU, with peak memory usage of only 402MB. Smaller models like Kitten Nano 0.8 have a higher error rate of 9.81%, making this model the most accurate among small models.

It is free for non-commercial use, while commercial use requires a separate license. It can be freely used in personal projects, research, education, and non-profit organizations. Crediting Underdog Woof or underdog.ai is a mandatory condition. Using this model's audio for model training or fine-tuning is prohibited.