Speech mode uses your browser's voice in/out. Replies are queued to the cluster's inference lanes via the gated path — first token can take a few seconds.