Whisper Model Quantized
The repository contains the whisper model quantized using the Smooth Quant using ONNXRuntime
- Only the Whisper decoder is quantized in the model
- The model has been modified to accept fixed shapes input of (1, 80, 3000) for the encoder and (1,448) for decoder.
- For inference the un-quantized encoder model and quantized decoder model is used.
- This model is for testing and could be modified in the future with better versions.
Evaluation:
The model achieves WER of 6.02%
on the librispeech_asr (clean)
test dataset