A developer posted an early preview of Alkove, a voice front-end for local assistants made of five small on-device models totaling about 7.5 MB, with code and weights under Apache-2.0. Lex turns a typed word into likely pronunciations, Spot listens for them in seven languages, Timbre builds voiceprints, Focus keeps only the user's voice, and Cue detects speech every 30 ms. The models run on a custom Rust inference engine compiled to WebAssembly and natively for Python and Android, with packages for npm, pip, Android Kotlin, and plugins for LiveKit Agents and Pipecat. The author reports that on AMI headset-mic meetings, other speakers' words reaching the transcript fall from about 17 per minute to 0.5 with Focus and a voiceprint, and that Spot recalls 67% of typed English words at one false alarm per hour, 90% in French and 85% in Portuguese; these figures are self-reported and not independently verified.