whisper.cpp
whisper.cpp copied to clipboard
Android run demo by small model
I run the demo on a real Android machine and copied the samll model, but the conversion time is too long. Is it a machine problem?
I have the same problem, Samsung s10e exynos, base data model
Make sure to use release build and select tiny or base model. On my Redmi note 9s (Cortex-A75), 48sec of voice sample can be transcribed in 37sec with base model.
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
Make sure to use release build and select tiny or base model.
This. There were a huge difference in audio conversion times between a debug apk and a release one.
Model | Variant | Input audio time | Time for ranscribing |
---|---|---|---|
Tiny ( ggml-tiny.en.bin ) | Debug | 11001 ms (11 sec) | 37519 ms (37.51 sec) |
Release | 11001 ms (11 sec) | 5123 ms (5.12 sec ) |
Using a quantized model had almost no conversion time difference , but a major advantage in apk size , without compromising in Accuracy (in my observation)
Model | Variant | Input audio time | Time for ranscribing |
---|---|---|---|
Tiny Q4_2 ( ggml-tiny.en-q4_2.bin ) | Debug | 11001 ms (11 sec) | 46517 ms (46.51 sec) |
Release | 11001 ms (11 sec) | 5309 ms (5.3 sec ) |
Also , clicking the benchmark button in android crashes the app. Don't know the reason since im not much familiar with android.