StudioAuthor launches browser TTS tool with Paradee optimized to 201× the Kokoro baseline on one CPU thread
Original title: Making Kokoro TTS 50-100x faster without knowing how
The author released a free in-browser TTS generator built on Paradee-8M v1.1, distilled from Kokoro-82M, and optimized it through 11 steps into a v2g model (2.32 MB) running on a custom 136 kB WebAssembly engine called pdn, reaching 201× the baseline on one CPU thread with first audio in about 10-60 ms, memory down from 646 MB to 100 MB, and all text processed locally in the tab. Quality held up: UTMOS rose from 4.39 to 4.45 and WER fell from 5.65% to 5.48%, and in one measured run a desktop CPU core on pdn (184×) beat an RTX 4070 Ti via onnxruntime-web's WebGPU backend (113×). All models and the engine are Apache-2.0, and v2f already serves as the production read-aloud voice of a web-novel site.
The author documents all 11 optimization steps from v1.1 to v2g with speed and size data plus quality metrics, offering a referenceable method and reproducible results.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.