A record of my research plans and experiences at the SJTU Audiocc Lab

Preface

It has been half a year since I joined the lab. Because I slacked off during the first semester of my sophomore year, I put my research tasks on hold, delayed a paper submission, and am even now still at the beginner stage of getting started, which is genuinely embarrassing. On March 13 I had a long talk with Prof. lcd and decided to adjust my research plan. The overall plan is to first get familiar with using espnet, and then, based on an idea, start preparing an ICASSP submission for the deadline in early September.

Research Plan

Getting familiar with espnet

Today Prof. lcd walked me through the general workflow of using espnet, introduced the overall structure and framework, and showed me how to do remote debugging in VS Code. He said I could think about what ideas might be worth submitting to ICASSP in September (it is probably still too early; they haven’t started preparing either).

So my task for the next week or so is to run the whole enh pipeline end to end on wsj0, from stage1 to around stage11, and to hand-write a separator (a transformer is recommended).

Note that stage1 does not need the MATLAB data-processing part; it can be removed with a few tweaks.

I’m setting myself a deadline: finish before next Wednesday, i.e. March 19. In the meantime I can spend more evenings at room 3-520 in the SEIEE building, listen to the senior students’ discussions, and learn more about this direction.

Below is a daily log of my research effort.

3-17

First day working in the lab, but there seemed to be hardly anyone around in the evening… At first I couldn’t quite remember what I was supposed to do; after poking around for a while I got some idea. See Running espnet speech separation from scratch for details.

Worked for almost three hours, stuck on setting up the environment. Ran into a bug that felt like an environment problem, but my advisor hasn’t replied yet. Sad.

Unbelievably, I was the only one left before ten o’clock. The PhD students’ lives must be quite full.

Currently stuck at stage1; data preparation is not yet done…

Later installed things manually, got it running successfully, and finished the first four stages of data preparation.

3-19

Spent an afternoon in the lab and finally resolved the environment issues completely. Went from espnet to espnet3, broke two environments along the way, and successfully finished stage5. Tomorrow or the day after I’ll start debugging the training part.

Translated from the Chinese original.