Voice conversation test

Press Talk, speak, press it again. Your words go to Whisper, the text goes to the language model, and the reply is spoken in the voice you pick. Every delay is measured from the moment you press stop. Whisper, the speech models and the language model all run on the GPU box at home; nothing leaves the house.

Reply voice

Speech model Qwen3-TTS Base, streaming VoxCPM2, sentence by sentence

Changing the speech model reloads the GPU box (about 15 to 35 s).

Shorter pieces: also break at commas and dashes

Earlier first sound, mostly for VoxCPM2; the phrasing can suffer.

Language model

Persona

Loading models…

The browser asks once for the microphone. The first reply after a pause can be slow while the language model loads; the page asks it to load as soon as you press Talk.

Conversation

Nothing said yet.

Voice conversation test

Conversation Start over

Conversation