**OpenASR** is a local-first speech-to-text engine written in Rust that runs transcription entirely on the user's own machine. Rather than shipping a single proprietary model, it acts as a unified runtime over **30 models across 16 families** — Whisper, Qwen3-ASR, Parakeet, SenseVoice, FireRed, Dolphin, Moonshine, Granite Speech, and Fun-ASR-Nano among them — all reachable through one binary on CPU and Apple Metal, with CUDA, Vulkan, and ROCm builds distributed separately. The project's defining choice is its **fail-closed privacy posture**. In the default local mode audio never leaves the device; remote compute exists only after an explicit pairing step, and there is no silent network fallback. When the engine cannot produce a transcript it reports why instead of quietly degrading to a cloud call. Every model download is checked against a **signed catalog** before it is allowed to run, which addresses a supply-chain gap most local-inference tooling leaves open — model weights are usually pulled from a hub with no verification beyond a hash the same hub supplies. Functionally it covers the practical transcription surface: batch conversion of files or whole folders into plain text, SRT/VTT subtitles, or JSON with word-level timestamps; live microphone captions with streaming partial results; system-audio capture for meetings and lectures; speaker diarization; and translation to English in a single pass. A `openasr serve` command exposes an OpenAI-compatible `/v1/audio/transcriptions` endpoint on localhost, so existing OpenAI SDK code can be repointed at a local base URL without modification. The repository is the Apache-2.0 open core behind a desktop application for macOS (Apple Silicon) and Windows x64, which wraps the same engine in a native GUI. Model packs each carry their own upstream license, recorded in the registry and pack metadata rather than blanketed under the engine's terms. The project is explicitly pre-v1 — the maintainer warns that CLI flags, API surface, and pack format may change between 0.x releases — with v0.1.30 shipping on 2026-08-05 during a period of near-daily tagged releases.