Voicet debuts ultrafast local speech-to-text

This is a serious latency demo first and a polished consumer app second, which is fine: local speech-to-text only becomes compelling when it feels instant.

• –The engineering story is the real hook: Rust, Candle, CUDA, and a vendored flash-attn fork keep the runtime much smaller than the Python/Transformers stack.
• –The workflow looks genuinely useful for power users: live transcription, type-into-any-app mode, hotkey pause/resume, automatic paragraph breaks, and offline WAV transcription.
• –The author's performance claims are compelling but self-reported: the repo cites 5x realtime on RTX 5080 and much faster startup than Python, while the Reddit post says it beats Mistral's demo and Speechmatics.
• –It is CUDA-only and hardware-gated today: RTX 3000+ with 11GB+ VRAM, Windows/Linux support, and DGX Spark still untested.
• –Using the full BF16 model helps explain the speed/quality story, but it also explains the GPU requirement and power draw.