Transcribe.cpp – ggml speech-to-text inference engine (github.com)

Transcribe.cpp has unveiled a powerful C/C++ speech-to-text (STT) inference engine, designed to run diverse STT model families efficiently through the GGUF models on the ggml runtime. The engine supports multiple GPU backends—Metal, Vulkan, and CUDA—enabling rapid inference, and efficiently manages CPU paths using tinyBLAS acceleration. With 16 model families and over 60 variants, Transcribe.cpp is significant for the AI/ML community as it delivers high-fidelity transcription capabilities accessible to a wide array of applications, all verified by Word Error Rate (WER) checks against reference implementations. This initiative not only emphasizes ease of integration across various platforms but also invests in accuracy through long-running WER checks, ensuring reliable output in critical applications. Each supported model, documented on Hugging Face under the handy-computer project, backs efficient reference implementations and rapid model conversion through user-friendly commands. The engine's technical prowess is backed by contributions from Mozilla AI's BiR Program and Modal for GPU credits, cementing its evolution into a production-grade engine for versatile transcription needs. The support from Blacksmith further guarantees smooth continuous integration, paving the way for ongoing improvements and updates.