Is it possible to run Whisper cpp with batch of wav files ?

[amitli1]

We can run Whisper cpp with one file.
Is it possible to run Whisper cpp with batch of wav files ?

[misutoneko]

Of course, it's quite handy with libfvad for example. Never used --help huh?
I think the context is lost between the files (not sure about that though - please correct me if that's wrong).

[amitli1]

I used help, but didn't saw it is possible.

What is the command to run batch of wav files ?

For example, when using batch with size 1 (single file), I'm using the following command:
./main -m models/ggml-base.bin -l iw -bs 8 -f example1.wav

It seems that -f take only 1 file.

[misutoneko]

Yes I see what you mean, it's not very clear. You can simply use multiple -f definitions, like so:

./main -m models/ggml-base.bin -l iw -bs 8 -f example1.wav -f example2.wav

[amitli1]

I tried and it seems that it run the files in sequential and not parallel (as batch works in deep learning).
Have you tried and got shorter running times ?

[misutoneko]

AFAIK while there is parallel processing (the -t switch) in whisper.cpp, it isn't implemented for multiple files. The same thing can be accomplished with an external setup that launches several instances of the main binary (xargs -P maybe?). Although you have to be careful with memory usage with that kind of setup.

In my own scripts I have a libfvad preprocess stage which generates a bunch of (generally very small) wav files, so parallel processing would be a good fit. The same thing could be done by whisper.cpp itself and it might give some more performance when the hw has lots of threads.
The running times are currently not an issue for me, so I haven't really pursued performance enhancements in this way.