IPD-Project: FPGA-Accelerated Automatic Speech Recognition, for TU München
DAS FINDEST DU FÜR DICH BEI UNS - That's in for you
DAS PROJEKT - About the Project
"Attention is all you need" - Transformer-based neural networks were only developed in 2018, but already conquered many deep learning domains. They constitute the state-of-the-art in image recognition, language modeling and automatic speech recognition (ASR).
Transformers have good performance, but their complexity rises quadratically with the sequence length. This poses challenges in embedded and battery-driven applications. To accelerate the network, we want to use a high-performance embedded hardware accelerators as computing device. Your task is to convert a Pytorch speech recognition network to the accelerator. This work is hands-on with embedded hardware.
Accompanying to the IDP, we recommend a suitable lecture at the faculty of electrical engineering of the TUM. This work can be done alone or in a group of up to three students.
DAMIT UNTERSTÜTZT DU UNS - Your Tasks
- Start with a pre-trained state-of-the ASR model
- Compile, and quantize the network
- Evaluate the performance based on inference speed and memory consumption
DAS BRINGST DU FÜR UNS MIT - Your Qualifications
- Experience with Linux, Python, Pytorch and Neural Networks (Deep learning lectures are OK)
- Experience with embedded systems or be interested in learning it
- Motivation to learn something new
- At least B2 German proficiency