wtf is cohere-transcribe-ruby?

This is a Ruby gem for transcribing Arabic and English audio into text using Cohere's speech recognition model. It runs entirely in Ruby with native C and C++ extensions, without requiring a Python runtime, a separate worker process, or a model server. The project mirrors the Python package's interface while keeping all inference inside Ruby. The core feature set covers the pinned Cohere Arabic and English model, compatible fine-tunes from Hugging Face, and local checkpoints. It includes audio decoding at 16 kHz, voice activity detection using Silero or an energy-based detector, fixed-window segmentation, batched inference, word-level alignment using a pinned MMS model, and subtitle generation. It also supports progress callbacks, reusable model sessions, and transactional file output so partial batches can be rolled back cleanly. Audio decoding works through a native adapter that binds FFmpeg's shared libraries directly, without launching the ffmpeg executable. Supported formats include AAC, FLAC, MP3, MP4, Ogg, Opus, WAV, WebM, and WMA. If the native adapter is missing, the gem can fall back to libsndfile and libsamplerate for a smaller set of formats. Decoding has monotonic deadlines, 30 seconds for duration probes and one hour for full decodes, and supports cooperative cancellation. Installation requires 64-bit Ruby 4.x on Linux or macOS. Windows is not currently supported. CPU builds are the default. CUDA builds need the NVIDIA toolkit, and Metal builds need Apple's Metal toolchain. The gem builds its native extension locally during installation, so you also need CMake and a C++20 compiler. A built-in doctor command checks the runtime and model access before you start. The Ruby API offers a simple one-shot method for single files and a persistent transcriber object for repeated calls. You can pass individual files, directories, or arrays, and the run returns successful, failed, and skipped results so one bad file does not abort a batch. Optional publication settings write TXT, SRT, VTT, and JSON files. The full README is longer than what was shown.