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.
github: 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.