Whisper Demo App

This app runs the Whisper model with the Qualcomm backend in ExecuTorch. The model itself runs on the Qualcomm HTP (Hexagon Tensor Processor), an AI accelerator in Qualcomm SoCs. You need an Android smartphone with a supported Qualcomm chipset.

Build the ExecuTorch Android library

Build the ExecuTorch Android library with QNN backend.

Export the audio processing and model .pte files

There are two steps, audio processing and the Whisper model (encoder+decoder), which are both done via ExecuTorch.

Run the script extension/audio/mel_spectrogram.py to export whisper_preprocess.pte
Run the script examples/qualcomm/oss_scripts/whisper/whisper.py to export whisper_qnn_16a8w.pte

Move these two .pte files along with tokenizer.json to /data/local/tmp/whisper on device.

Run app

Open WhisperApp in Android studio. Copy the Android library executorch.aar which should have audio JNI bindings and the Qualcomm HTP libraries, into app/libs.

20250822_125343.mov