
Speech Processing & Machine Learning Lab @ntu_spml | National Taiwan University
Joined August 2009
- Tweets35
- Following86
- Followers30
- Likes593
Unsupervised models that learn from raw speech have been limited to discovering acoustic/lexical information, such as phonemes or words.
Can we build unsupervised speech models that learn high-level linguistic information like syntax?
#ICASSP2023
Another interesting discovery I wish to highlight is that Approach 2, when trained without any textual data, still shows signs of learning the different branching directions (a lang.-specific syntactic property) of English and Korean from speech directly.
Code, Video & Poster: github.com/roger-tseng/speec…
Full paper: arxiv.org/abs/2303.08809
If you find this topic interesting, come chat!
I'll present our paper tomorrow from 3:40 PM - 5:10 PM at the Audio and Text Segmentation, Tagging and Parsing (SLT-P9) poster session at #ICASSP2023
Roger Tseng retweeted
The research topic I've been working on for a long time is going to be incorporated into iOS 17 by Apple (of course, I didn't write the code in iOS): 9to5mac.com/2023/05/16/apple…
Roger Tseng retweeted
``Analysing Discrete Self Supervised Speech Representation for Spoken Language Modeling. (arXiv:2301.00591v1 [cs.CL]),'' Amitay Sicherman, Yossi Adi, ift.tt/KwI4mGs
Roger Tseng retweeted
Abdelrahman Mohamed (Meta), Shinji Watanabe (CMU), Tara Sainath (Google), Karen Livescu (TTIC), Shang-Wen Li (Meta), Shu-wen Yang (NTU), Katrin Kirchhoff (Amazon), and I will give a tutorial about self-supervised learning for speech at NAACL 2022. 2022.naacl.org/program/tutor…
Note to self to check this out later. What does this suggest for unsupervised language acquisition?
ArXiv arxiv.org/abs/2207.02098: Empirical study on NNs for learning formal languages. RNNs and Transformers fail to generalize on non-regular tasks, LSTMs solve regular and counter-language tasks, and only NNs with structured memory learn context-free and context-sensitive tasks.
1
Roger Tseng retweeted
"Progressive Stage-wise Learning for Unsupervised Feature Representation Enhancement"
Someone finally beat regular backpropagation on a realistic problem—specifically, in the context of self-supervised learning (SSL). [1/12]
Note to self to check this out later. What does this mean for unsupervised language acquisition?
ArXiv arxiv.org/abs/2207.02098: Empirical study on NNs for learning formal languages. RNNs and Transformers fail to generalize on non-regular tasks, LSTMs solve regular and counter-language tasks, and only NNs with structured memory learn context-free and context-sensitive tasks.
1
Roger Tseng retweeted
Great advice by @EugeneVinitsky
-- Run the simplest experiment first
-- Write a mini-version of your paper before getting started
-- Keep an ultra-detailed notebook
-- Every paper should be reproducible via 1 script
-- Save all your videos in one place
-- Present regularly
A lot of blog posts on research vision, less so on the day to day of managing your research, experiments, and productivity. I’ve repeated this to students enough times so I’ve written it up tinyurl.com/Research-product…
Roger Tseng retweeted
``Self-supervised Learning with Random-projection Quantizer for Speech Recognition. (arXiv:2202.01855v2 [cs.CL] UPDATED),'' Chung-Cheng Chiu, James Qin, Yu Zhang, Jiahui Yu, Yonghui Wu, ift.tt/N1kKMaI
Roger Tseng retweeted
A very interesting read! There's indeed a huge gap between usable tech a musician would enjoy and "fancy" tech the AI community finds worth investigating!🧐
On the Development and Practice of AI Technology for Contemporary Popular Music Production: transactions.ismir.net/artic…
Roger Tseng retweeted
There are good tutorials on the concepts behind transformers out there. Also, there tons of tutorials on using the HF transformers lib.
But how about a step-by-step guide on coding BERT, GPT, BART from the ground up? Here's @DanielPressel's amazing repo: github.com/dpressel/mint
Nice checklist for training neural nets better karpathy.github.io/2019/04/2…
1
Roger Tseng retweeted
Meet BasicPitch.io — a 🎹🎸🎻🪕🎺🎷🪗 🎤-to-MIDI converter that uses about 70 billion fewer parameters than those fancy neural networks. engineering.atspotify.com/me…
#MachineLearning
Good read. I would benefit a lot from more experience in quickly iterating between new solutions forecastegy.com/posts/is-com…
1













