Georgi Gerganov

My Raspberry Pi 4 reports `armv7l`, but I can use `-mfpu=neon-fp-armv8` successfully. It has much better performance compared to the proposed `-mfpu=neon`. Does `-mfpu=neon-fp-armv8` not work for you?

Yes, it looks like I am using a 32-bit OS on my 64-bit RPi4. Thanks for clearing this up. It would be very useful to have a branch in the...

``` $ cat /proc/cpuinfo |grep 'model name' model name : ARMv7 Processor rev 3 (v7l) ```

The current diarization approach only kind-of works with stereo audio. It does not work if you convert mono audio to stereo audio.

Don't think you can do anything at the moment to improve the performance. In the future, quantised models might be useful for such use-cases, so keep track of progress in...

In addition to @tinoue comment, in the future you might be able to use quantised models for better performance on mobile devices - follow #540 for more informatiom

`whisper.cpp` has a built-in matrix-multiplication routine. Additionally, it supports using a 3rd party matrix multiplication provided by the OpenBLAS library. Which one is faster depends on the CPU and hardware...

See if the "Fine-tuned models" section has any useful info: https://github.com/ggerganov/whisper.cpp/tree/master/models#fine-tuned-models

Try adding `--no-fallback` parameter

@regstuff I haven't done accuracy evaluation yet. The accuracy definitely drops, especially for the smaller models. But I cannot say how much yet. For example, I observe that GPT-2 `117M`...