Improve performance, inspire from whisper-jax #1773
Unanswered
thewh1teagle
asked this question in
Q&A
Replies: 1 comment
-
Whisper-jax segments audio into 30-second chunks and processes them in batches, which enhances the Compute-to-Memory Ratio. You can find more about this ratio here: Compute-To-Memory Ratio. In most cases, especially when using any kind of accelerator, the main limitation in decoding is the memory bandwidth. We can definitely implement this feature. It just might take a little time since I'm not a full-time developer here. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I just saw the project
whisper-jax
which says it works 70x faster than original usinggoogle-jax
.But they use
Python
and it's not really usable at this moment for using in projects like vibeBut 70x faster is really attractive.
Do we have a way to improve
whisper.cpp
to get similar performance? what so special ingoogle-jax
?Beta Was this translation helpful? Give feedback.
All reactions