Voice Bench

Expand numbers, dates, times and abbreviations before synthesis

Sent to the model

Respellings applied before synthesis, whole words only, case-insensitive. Works on both engines. Kokoro's phoneme layer isn't reachable from JavaScript, so respelling is the practical route.

Written	Spoken as

Add a word

Generate sentence by sentence, with per-sentence pacing

Kokoro concatenates the chunks into one file. Web Speech queues them as separate utterances.

Apply the processing chain to Kokoro output

Rendered offline, so the download is the processed audio. Web Speech can't be processed; the browser never hands over the samples.

Match loudness across takes Reset to flat