wtf is moss-transcribe.cpp?

moss-transcribe.cpp is a from-scratch rewrite of an AI model called MOSS-Transcribe-Diarize, built by the LocalAI project. The original model transcribes audio with multiple speakers, figures out which words belong to which speaker, called diarization, and marks exactly when each part of the speech happened, all in a single pass rather than stitching together separate tools for each task. This port reimplements the whole thing in C++ using a library called ggml, so it can run without needing Python, PyTorch, or a CUDA toolkit at the moment you actually transcribe audio, and the entire model fits in one self-contained file format called GGUF. What makes this project notable is that it claims to be exactly as accurate as the original: every piece of the model is checked against the real PyTorch version and matches almost perfectly, so the text it produces is identical, not just similar. On a regular CPU it also runs noticeably faster than the original, around 1.6 to 1.8 times faster on the same hardware, while using less memory, and it can shrink the model down to a fraction of its original size through a technique called quantization, again without changing the output text. It also supports running on GPUs through several backends, including CUDA, Metal, Vulkan, and AMD's HIP, and the CUDA path has been verified to give the identical result on modern NVIDIA hardware. The output looks like a script with timestamps and speaker labels attached directly to the words, for example marking where one speaker's line ends and another begins along with the exact second it happened. Long recordings are handled by splitting the audio into thirty second chunks, processing each one, and stitching the pieces back together while keeping the timing consistent across the whole recording. To use it, you build the C++ project with CMake, then convert an official MOSS-Transcribe-Diarize model download into the GGUF format using an included Python script, which is only needed once during conversion and not while actually transcribing audio. The project is aimed at people who want fast, accurate, multi-speaker transcription that can run entirely on their own hardware without relying on Python or a GPU.