Runs your mic through the current Sound settings live. Use headphones — otherwise it will feed back.
Monitor path is a straight wire while every effect slider sits at zero.
Longer text is split at punctuation marks and rendered piece by piece, so the first sound comes sooner. The Stop button above cancels everything.
Translation goes from the recognition language to the target language. Through the proxy both providers work reliably; without the proxy, Google can fail on CORS.
The local engine runs on the device instead of Google's servers: no round trip, works offline, and it does not silently give up after long pauses. Needs Chrome 139+ (ChromeOS ships it).
When you go quiet this long, the recognizer is told to finalise — it then transcribes the whole buffered sentence properly instead of the app grabbing the rough live guess.
One name per line. These are handed to the recognizer as contextual hints, so it actively expects them — and anything that comes back close to one of them gets snapped to the correct spelling.
One rule per line in the format wrong = right. Applied before reading aloud, whole words, case-insensitive. Use ✎ on any line in the log to teach a new rule in one click.
Pre-filled from the KEYS constant at the top of the script. What you type here overrides it and is kept in localStorage.
Recreated characters, not original recordings. Pitch, tempo, formant filter, and tremolo are set to approximate them.