transcribe.cpp
A C/C++ speech-to-text inference library that runs 16 model families and more than 60 model variants on the ggml runtime, with Metal, Vulkan and CUDA GPU backends plus an accelerated CPU path.
transcribe.cpp is a C/C++ speech-to-text inference library. It loads GGUF-format models and runs them on the ggml runtime, with Metal, Vulkan and CUDA backends for GPU inference and a tinyBLAS-accelerated CPU path, supporting both streaming and batch transcription.
It covers 16 model families and more than 60 model variants, including Canary, Canary-Qwen, Cohere Transcribe, Fun-ASR-Nano, GigaAM-v3, Granite Speech, MedASR, Moonshine and a streaming Moonshine variant, MOSS-Transcribe-Diarize, Multitalker Parakeet Streaming and Nemotron speech models, with capabilities that vary by model such as translation, diarization, streaming and token or word timestamps. Every model published under the project is numerically verified and tested for word error rate against its reference implementation.
It is written in C++ and licensed under the MIT license.