Transcribe.cpp(workshop.cjpais.com)
696 points by sebjones 20 hours ago | 144 comments
tl;dr: Transcribe.cpp is a new ggml-based transcription library from the maintainer of Handy, supporting 16+ ASR model families with GPU acceleration via Vulkan, Metal, CUDA, and TinyBLAS across Mac, Windows, and Linux. It's numerically validated and WER-tested against reference implementations, functions as a mostly drop-in whisper.cpp replacement, and ships with first-party bindings for Python, JS/TS, Rust, and ObjC/Swift. The author built it to solve the pain of distributing cross-platform local ASR, arguing current options (whisper.cpp, ONNX, MLX) leave performance and model coverage on the table.
HN Discussion:
  • Praise for the rigor and ambition of a solo-developed contribution to the ASR ecosystem
  • Validation of the article's premise that swapping ASR models is painful and this library helps
  • Pushback on the article's characterization of ONNX as CPU-only, noting it supports GPU/TensorRT
  • Requests for adjacent features not covered (IPA phonetic transcription, video/system audio, continuous dictation)
  • Appreciation for numerical validation and WER testing as a differentiator from typical 'supports everything' claims