amd/NPU-Whisper-Base-Small

AMD Python NPU Whisper-Small Script

This directory contains a sample python script to run the Whisper-Base-Small model on CPU and NPU. A single sample of the LibriSpeech dataset has been taken and converted from flac to wav (test.wav). Both encoder and decoder models have been converted to ONNX using static tensor shapes. The sequence length is 448 tokens.

The paths to the models, input file, and other needed constants are defined in upper case at the top of the file. The defaults are set to run the existing file test.wav

The sample text produced by the script should correspond to the first line of the transcribe.txt file

Note: The models need to be compiled and cached. This will happen the first time they are run. It may take several minutes for the compilation to complete. Compilation only needs to be done once.

Instructions for running Whisper-Small

Open Developer Command Prompt for VS 2022 Activate the Ryzen AI 1.5 Conda env at the command prompt

conda activate ryzen-ai-1.5.0

If you don't have the requirements listed in requirements.txt , run

pip install -r requirements.txt

To run on CPU

At the command prompt, enter

python -m amd_whisper 

To run on NPU

At the command prompt, enter

python -m amd_whisper --npu