Show HN: Optimized CUDA forced alignment kernels. 10x – 1000x faster

Category: ai-ml

Tags: cuda, ctc, forced-alignment, speech-recognition, gpu-acceleration

Score: 7.3/10 (Innovation: 7, Technical: 8, Documentation: 8, Utility: 6)

Fast Forced Aligner provides optimized CUDA kernels for forced alignment of CTC emissions to transcripts, achieving 10x-1000x speedups over existing CPU and GPU implementations. It offers both an optimized chunked algorithm and a conventional full dynamic programming mode behind a simple Python API, targeting high-throughput speech processing pipelines.

Target audience: ML engineers and speech recognition researchers

Repository: https://github.com/brthor/fast-forced-aligner · Python · MIT

View on Hacker News