Release notes, engineering deep-dives, and field reports on running speech recognition entirely on your own hardware.https://openasr.org/en-ushttps://openasr.org/blog/milestone-01x/https://openasr.org/blog/milestone-01x/From 0.1.0 to 0.1.14 in ten days — the foundation-laying sprint that cut memory in half, shipped 26 models, landed on eleven platform targets, and closed every idle-unload gap before moving on to 0.2.Tue, 14 Jul 2026 00:00:00 GMThttps://openasr.org/blog/windows-launch/https://openasr.org/blog/windows-launch/OpenASR Desktop launches on Windows today — the same local-first speech-to-text that runs on macOS, now on Windows 10+ x64. Same open core, same privacy boundary, same 26 models.Tue, 14 Jul 2026 00:00:00 GMThttps://openasr.org/blog/openasr-0-1-0/https://openasr.org/blog/openasr-0-1-0/The first public release of OpenASR: a local speech-to-text app for macOS, built on an Apache-2.0 open core. What works today, what doesn't yet, and where the code lives.Sat, 04 Jul 2026 00:00:00 GMThttps://openasr.org/blog/pure-rust-diarization/https://openasr.org/blog/pure-rust-diarization/OpenASR adds WhisperX-style diarization as pure-Rust stages around an untouched ASR core. Forward passes are bit-close to ONNX — a numerical check, not an accuracy claim.Sat, 06 Jun 2026 00:00:00 GMThttps://openasr.org/blog/signed-model-distribution/https://openasr.org/blog/signed-model-distribution/OpenASR splits model trust into a sha256-pinned pack and an Ed25519-signed catalog manifest, wired into a pull chain that fails closed on any mismatch.Thu, 04 Jun 2026 00:00:00 GMThttps://openasr.org/blog/batched-serve-decode/https://openasr.org/blog/batched-serve-decode/An opt-in, default-off batching path multiplies aggregate serve-mode decode throughput on a GPU, while the single-request path stays provably byte-identical to today's decode loop.Tue, 02 Jun 2026 00:00:00 GMThttps://openasr.org/blog/x86-gpu-device-resident-kv/https://openasr.org/blog/x86-gpu-device-resident-kv/On one AMD GPU running qwen3-asr-0.6b q8, profiling showed KV host-to-device copies dominated decode. A device-resident cache removed that bottleneck.Sat, 30 May 2026 00:00:00 GMThttps://openasr.org/blog/benchmarking-whisper-cpp/https://openasr.org/blog/benchmarking-whisper-cpp/OpenASR's bench suite measures a 7.9% edge over whisper.cpp on one host with one model. The spread is shown; the goal is regression detection, not a blanket claim.Wed, 27 May 2026 00:00:00 GMThttps://openasr.org/blog/inside-oasr-format/https://openasr.org/blog/inside-oasr-format/Why an OpenASR model is a single GGUF-backed .oasr file with its descriptors embedded as metadata - and why that file is a portable pack, not a publisher-signed artifact.Sat, 09 May 2026 00:00:00 GMThttps://openasr.org/blog/local-by-default/https://openasr.org/blog/local-by-default/OpenASR transcribes audio where it is recorded: an offline-default, no-auto-download boundary you can read in the source, all the way through to diarization.Thu, 30 Apr 2026 00:00:00 GMThttps://openasr.org/blog/openai-compatible-localhost/https://openasr.org/blog/openai-compatible-localhost/Point an OpenAI transcription client at localhost — OpenASR serves three endpoints and six response formats offline, with explicit gaps documented.Sat, 04 Apr 2026 00:00:00 GMThttps://openasr.org/blog/quantizing-accent-quality/https://openasr.org/blog/quantizing-accent-quality/What the OpenASR perf suite shows about quantization on one LibriSpeech clip: most levels hold WER 0.00%, speed tracks precision rather than compression, and a couple of tiny models are genuinely lossy.Sun, 22 Mar 2026 00:00:00 GMT