Useful Transformers is a library for efficient inference of Transformer models. The focus is on low cost, low energy processors to run inference at the edge. The initial implementation is aimed at running OpenAI’s Whisper speech-to-text model efficiently on the RK3588 processors’ based single-board computers. The tiny.en Whisper model runs transcribes speech at 30x real-time speeds, and 2x better than best known implementation.
Getting started
The easiest way to try out Whisper transcription is to install the release wheel package.
# Preferably inside a virtual environment
$ python -m pip install https://github.com/usefulsensors/useful-transformers/releases/download/0.1_rk3588/useful_transformers-0.1-cp310-cp310-linux_aarch64.whl
Try transcribing a wav file.
$ taskset -c 4-7 python -m useful_transformers.transcribe_wav
Ever tried, ever failed. No matter, try again. Fail again. Fail better.
Performance
The plot shows useful-transformers Whisper tiny.en model’s inference times across the examples with varying durations. useful-transformer is 2x faster than faster-whisper‘s int8 implementation. useful-transformer uses FP16 matrix multiplication on the NPU available in the RK3588 processor. The majority of benefit comes from the large matrix multiplications (of sizes 1500x384x384 for the tiny.en model) in the encoder.
TODO
- Whisper tiny.en
- Whisper base.en
- Larger Whisper models
- Use int8 matmuls from the librknnrt
- Use int4 matmuls (request Rockhip for int4 matmul kernel
