👀 Problem

Whisper emerges as a standout among the most cutting-edge DNN-based ASR systems currently available. These exceptional systems empower a wide array of human-computer interfaces, notably including the widely-used Alexa. The primary objective of this project is to delve into the effectiveness of quantisation as a method for efficiently reducing the inference cost associated with these models.

Introducing Whisper

https://github.com/openai/whisper

💭 Proposal

Base elements

The student will be tasked with implementing the Whisper network and its benchmarking flow within MASE, our team's internal compiler stack. Additionally, they will be required to utilize the various quantisation functions in MASE and assess their performance.

Extensions

• Mixed arithmetic quantisation in MASE and evaluate its performance on Whisper.

🛫 Plan

• Read and understand the different implementations of quantisations and how we have implemented in our in-house compiler MASE.
• Implement Whisper in MASE together with a few datasets to evaluate your implementation.
• Implement mixed-quantisation for Whisper manually to test functionality.
• Use the existing quantiisation search capability in MASE and evaluate its performance

🖥️ Skill Requirements

• A good understanding of Machine Learning