Ash holds a local conversation
Spoken questions now become local replies and clear speech, with three-turn context and Windows playback.
Ash completed a working spoken conversation loop: a USB microphone captured a question, Whisper transcribed it, a local Qwen model generated a reply, and Piper turned that reply into speech. A Windows launcher downloaded and played the reply before starting the next microphone turn.
Three-turn sessions demonstrated recall within a conversation. Ash remembered a name and a preference supplied earlier in the same session. Each new launch starts fresh; this is not persistent personal memory.
Ethernet also removed the terminal lag reported during setup. That made the development loop much easier to use. The speech playback was clear and audible through Windows speakers; the monitor speaker was still an open item.
The language model runs on the Raspberry Pi CPU. The installed Hailo accelerator is not running this model. A larger model was tested, but its slower replies made the smaller model the current development choice.
This milestone is deliberately bounded: three turns, sequential listening and playback, and no live file or sensor access through the conversation interface. Reply accuracy and natural conversation still need work. The public Command Center remains a separate animation, not a live connection to the Pi.