Small speech models often trade accuracy for portability, but Woof’s reported benchmark results challenge that assumption: its two-voice model records a 1.42% word error rate on 500 English hard-case sentences while fitting in a 40.6 MB download. The notable idea is not just a compact footprint; it is making local, CPU-only speech generation practical without ignoring the text forms that commonly trip TTS systems up.
Key Capabilities
- Built for tricky written forms: Rule-based normalization handles numbers, dates, times, money, codes, and symbols, while a lexicon helps with difficult names. This targets everyday inputs that can sound awkward when read literally.
- Two voices in one small model: Juno and Mira are bundled together; on the reported test, Mira scored 1.42% WER and Juno 1.53%. That gives users a choice without downloading separate voice packages.
- Local inference without a GPU: The NumPy runtime generates 24 kHz mono audio on a CPU. The project reports 13.8× real-time speed for Juno on one Apple M5 Max CPU thread, making offline use plausible on capable laptops.
- Strong reported results for its size: On the project’s Gradium TTS Hard Cases evaluation, Woof’s two-voice version scored 1.42% WER, compared with 1.89% for Fish Audio Drama 3 and 2.16% for Phonon 31m. These are the project’s own tests, so the ranking should be read in the context of its dataset, voices, and evaluation setup.
Who It’s For
A good fit for developers who need lightweight, offline English narration, especially for text containing dates, contact numbers, or codes. Look elsewhere if you need multilingual output, a broad voice catalog, commercial use without a separate license, or independently established benchmark comparisons: the license restricts commercial use and model/audio training, and the reported evaluation is limited to English.
