Alkove: 7.5MB on-device voice front-end keeps only your voice before STT
cocktailpartyfixer · reddit · 2026-10-21
While running production LLM voice assistants, the author found most trouble happens before the LLM: bystanders leak into transcripts and trigger false barge-ins. The answer is Alkove, an open-source (Apache-2.0) on-device voice front-end of five tiny models (7.5MB total):
- Lex (1.8MB): converts a typed wake word into likely pronunciations — wake words are just text, no training
- Spot (2.8MB): listens for those pronunciations in 7 languages
- Timbre (0.74MB): builds a voiceprint from three wake-word takes, keeps learning
- Focus (1.1MB): keeps only your voice, removing noise, music and other speakers before STT
- Cue (1.1MB): detects speech every 30ms; with the print, only your voice opens a turn
Tech: a custom Rust inference engine (not onnxruntime), bit-identical across WebAssembly, Python and Android; browser engine is 134KB brotli, fed by AudioWorklet workers, no COOP/COEP headers. Phone single-thread speed: Focus 38.6ms per second of audio. Ships via npm, pip, Android, plus LiveKit Agents and Pipecat plugins.
Benchmark: on AMI meetings, words from other people reaching the transcript drop from 17/min to 0.5 with Focus + voiceprint. Limits: near-mic single-user only, same-register overlapping voices still hard, 16kHz audio.
More from coding & agent
- Training an original-character LoRA for MiniMax H3: full workflow and pitfalls — DeliciousGorilla · 2026-10-21
- Dev builds open-source Chrome extension unifying Gemini and Antigravity on Windows — iamfivestars · 2026-10-21
- Dev interviews his Claude Code agents on a video standup — the human did worst — stickygoose · 2026-10-21
- MCP server SDK CVE-2026-111770: shared McpServer leaks responses across clients — Future_AGI · 2026-10-21
- When a Customer's AI Assistant Calls Support, Who Checks What It's Authorized to Do? — IrfanZahoor_950 · 2026-10-21
- Two weeks in SF: fundraising goes binary, top devs now review agents not code — ai · 2026-10-21