Fabio Angela · 2026 · Preprint

Teacher in the Loop

Kokoro-82M distilled into 8M-parameter single-voice students, in English and Italian. Type something below: the voice runs entirely in your browser.

Try it

Runs on your device
Ctrl + Enter  
Phonemes
Not exactly the Python package. The voices are the released int8 graphs, unchanged, but here the text is turned into phonemes by a JavaScript port of the pipeline the paper uses. Italian gives the same phonemes on all 2,000 test sentences. English, which has no part-of-speech tagger in the browser, gives the same phonemes on 95% of them; the rest differ in a word or two (mostly a noun/verb homograph or the direction of a quotation mark), so some sentences can sound slightly different from the titl package and from the paper's evaluation.

Everything happens in this tab. Text becomes phonemes with Misaki's lexicon and eSpeak NG compiled to WebAssembly; the voice's int8 ONNX graph (text side and decoder, 9.2 MB) then runs on ONNX Runtime Web, on one CPU thread. What you type is not sent anywhere. The first use downloads the runtime and the voice.

Listen to the paper's samples

The first five of the 200 held-out sentences of each voice, in their fixed order, not chosen by listening. Every system on the same sentence, side by side.

What you are hearing

Decoder alone. Every decoder receives the teacher's own durations, pitch, energy and phoneme features for the sentence and the same random excitation seed, so only the decoder differs. These clips are rows of the paper's 200-sentence evaluation; the UTMOS next to each system is its mean over those 200 sentences.

End to end. Phonemes in, waveform out, through the released int8 ONNX graph of each model (text side and decoder), as the titl package plays them. Paradee released only an English voice.

Sentences and audio. WikiText-103 for English, the Italian Wikipedia for Italian; no student saw them in training. 24 kHz, 16-bit FLAC, not normalized or edited. Headphones help: the differences between the stronger systems are small.

Tip: switching to another system while a clip plays keeps the position, for quick A/B listening.

1As of August 2015, there were 169 proposed targets for these goals and 304 proposed indicators to show compliance.

Decoder alone, on the teacher's features
End to end, from phonemes

2White dwarf stars, if they have a near companion, may then become Type Ia supernovae.

Decoder alone, on the teacher's features
End to end, from phonemes

3According to Herodotus, this was because they had decided in council that they could not beat the Allies in a naval battle.

Decoder alone, on the teacher's features
End to end, from phonemes

4The ashes were later taken to Veluvana, where they were enshrined in a silver stupa.

Decoder alone, on the teacher's features
End to end, from phonemes

5Ten priests of the Evangelical Lutheran Church of Finland and 90 (moderate) socialists were executed also.

Decoder alone, on the teacher's features
End to end, from phonemes