Here with you.
Hold the button or Space. Release for a squeak.
No speech model loaded.
A short audio file, processed here on your device. Only its first 15 seconds are used.
Try a happy sample A synthetic English voice.
Your microphone is opened only while holding Talk. It closes on release, tab hide, or cancellation. Audio and words stay in memory for the reply; they are never uploaded or saved. Preferences stay in this browser.
WebGPU accelerates the paper stage and supported inference. Unsupported model operations use the local CPU. Word-aware models download from Hugging Face only when requested. No browser speech-recognition service, paid API, or Mangatsika connection is used.
Emotion estimates can be wrong. Word-aware replies mirror one of seven broad emotions; this is a squeaking companion, with no generative conversation or factual answers. Quick-only mode hears delivery without understanding words.
Credits & model licenses · Reset preferences