yum-archive/TaSTT-Whisper

High-performance GPGPU inference of OpenAI's Whisper automatic speech recognition (ASR) model

git clone https://git.yummers.dev/yum-archive/TaSTT-Whisper

KonstantinMinor, documentation64d61ad

master
branchesmasterall refs →
1.7 KiB30 linesraw
1This folder contains 2 sample speech audio clips, `jfk.wav` and `columbia.wma`
2
3jfk.wav is from whisper.cpp repository.
4
5columbia.wma is from Wikipedia: https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg
6I re-encoded the audio from Ogg Vorbis into Windows Media Audio, because Media Foundation is unable to decode Vorbis.
7
8The rest of the text files in this folder are the outputs of the in-app performance profiler, when the app was transcribing these two audio clips on three different computers.
9
10The “1080ti” files are from my desktop, which has nVidia GeForce 1080Ti GPU.
11
12The “vega8” files are from the same desktop, when using the GPU integrated into AMD Ryzen 7 5700G processor.
13
14The “vega7” files are from my laptop, the GPU is integrated into AMD Ryzen 5 5600U processor.
15The laptop model is HP ProBook 445 G8. While running the tests, the laptop was on battery power.
16
17The “1650” files are from another desktop with nVidia GeForce 1650.
18
19The file names with “medium” in the middle were made with “ggml-medium.bin” Whisper model, with “large” were made with “ggml-large.bin” model.
20
21In theory, 1080ti delivers 10.6 TFlops FP32 and 484.4 GB/second VRAM bandwidth.
22
23The AMD APU in that desktop delivers 2.0 TFlops FP32, and 53.3 GB/second memory bandwidth.
24
25That variant of 1650 delivers 2.6 TFlops FP32, and 128.1 GB/second VRAM bandwidth.
26
27The AMD APU in the laptop delivers 1.6 TFlops FP32, and 51.2 GB/second memory bandwidth.
28
29The loading times in these logs make little sense because they depend heavily on the disk cache implemented in the OS kernel.
30Ideally, I needed to reboot Windows before every test, but I didn’t do that.