warpt-ctc

warp-ctc is an open-source, high-performance parallel implementation of the Connectionist Temporal Classification (CTC) loss, supporting both CPU and GPU (CUDA) execution. Designed for deep learning practitioners and researchers, it enables efficient and scalable training of models—particularly recurrent neural networks—on sequence data without requiring pre-aligned input-label pairs. warp-ctc comes with bindings for Torch and a simple C interface, making it a flexible choice for integrating CTC loss into end-to-end pipelines like speech recognition.