Teacher in the Loop: 8M-parameter Kokoro voices in English and Italian, 9 MB int8 ONNX, running in the browser

MithrilMan released Teacher in the Loop, a set of 8M-parameter single-voice text-to-speech models distilled from Kokoro-82M in English and Italian, packaged as 9 MB int8 ONNX files that run entirely in the browser via ONNX Runtime Web and eSpeak NG compiled to WebAssembly. The project, installable with `pip install teacher-in-the-loop`, follows up on Paradee: re-running Paradee's published training scripts produced decoders scoring UTMOS 1.94 and 2.89 against 4.37 for the released decoder, which the author attributes to Kokoro's decoder normalizing over the whole utterance and drawing a random excitation, making the 1.6 s crop target unpredictable. The fix keeps the frozen teacher in the loop, rendering each crop with the student's excitation while the student also matches the teacher's iSTFT head, without changing the architecture.

Hi all, I’d like to share a project and ask for feedback: small single-voice TTS models distilled from Kokoro-82M, in English and Italian. Try it: Teacher in the Loop - a Hugging Face Space by MithrilMan https://huggingface.co/spaces/MithrilMan/teacher-in-the-loop It runs entirely in your browser ONNX Runtime Web, eSpeak NG compiled to WebAssembly . The Listen section plays the teacher, Paradee’s released decoder, the published recipe and mine on the same held-out sentences. What it is af heart English , if sara and im nicola Italian pip install teacher-in-the-loop , then titl --voice if sara "Buongiorno a tutti." -o out.wav Why it exists It’s a follow-up to Paradee. Re-running Paradee’s published training scripts, my decoders ended at UTMOS 1.94 and 2.89, against 4.37 for the released one. The target of the first decoder stage is a 1.6 s crop of a teacher render of the whole sentence. Kokoro’s decoder normalizes over the whole utterance and draws a random excitation, so that target can’t be predicted from the crop. Putting the frozen teacher in the loop fixes it without changing the architecture: the teacher renders each crop with the student’s excitation, and the student also matches the teacher’s iSTFT head.