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

Developer Bryan Thornbury released optimized CUDA forced-alignment kernels via GitHub (brthor/fast-forced-aligner) and a blog post, reporting speeds 10x to 1000x faster than popular open-source options, including the torchaudio implementations he previously used. Thornbury built the kernels to accelerate his own forced-alignment pipelines using Astra and published them for others to use.

blog post: https://www.bryanthornbury.com/posts/fast-forced-alignment https://www.bryanthornbury.com/posts/fast-forced-alignment github: https://github.com/brthor/fast-forced-aligner https://github.com/brthor/fast-forced-aligner I developed optimized CUDA kernels to speed up my forced alignment pipelines using Astra. To my surprise, I was able to achieve speeds 10x - 1000x faster than popular existing open source options, including the torchaudio implementations I was using before. I'm sharing them for anyone else who finds them useful. The acceleration from AI is getting wild. Comments URL: https://news.ycombinator.com/item?id=50310134 https://news.ycombinator.com/item?id=50310134 Points: 1 Comments: 0