Georgi Gerganov

Maybe @fitzsim can provide some info - they developed most of the PPC64 port

Yes - you are correct. It seems I didn't design the `whisper_context` in the best possible way. For now, you can achieve parallel processing using a single `whisper_context` by hacking...

Try to update `node`. It works with `v18.12.1`

Yes, I think the tokens are as follows: - token 0: `*` - token 1: `wind` - token 2: ` blowing`

You are correct that having more context in general improves the quality of the ASR quality. However, more text context makes the decoder slower, so I thought it is OK...

Reopen if issue persists

I believe this has been resolved at some point. Feel free to reopen if problem persists

Sure, you can install it and manually modify the Makefile with the correct path. In the future, we might add option to specify the path via environment variable.

OpenBLAS does not help on Mac - `whisper.cpp` already uses [Apple's Accelerate BLAS](https://developer.apple.com/documentation/accelerate/blas?language=objc) which is superior.

I think we can measure and report both. On non-Windows system we can use the existing `ggml_cycles()`: https://github.com/ggerganov/whisper.cpp/blob/d1ea1220ffeff846efd0578e8df3b1d1b802bb5d/ggml.c#L315-L321