Transcribe.cpp offers a whisper.cpp drop-in with 60-plus ASR models

Sources: Transcribe.cpp project page, HN discussion
Summary: Transcribe.cpp is an offline speech-to-text inference library presented as a near drop-in replacement for whisper.cpp that keeps compatibility with existing .bin model files while extending support to 16 ASR model families and more than 60 models. It adds acceleration backends for Vulkan, Metal, CUDA, and TinyBLAS, and the author states each model is numerically validated and word-error-rate tested against a reference implementation before release. The author reports faster-than-real-time transcription with state-of-the-art models on low-power hardware, citing an RK3566 SoC.
Comments: HN commenters pair it with the same-day Moonshine Micro release as a local-inference theme, and one asks why Metal runs roughly ten times faster than Vulkan in the author's numbers.
Why it matters: Broadening whisper.cpp-compatible tooling to many ASR families and GPU backends lowers the cost of embedding local transcription across platforms.