Timezone: »

Self-Supervised Learning for Speech and Audio Processing
Abdelrahman Mohamed · Hung-yi Lee · Shinji Watanabe · Shang-Wen Li · Tara Sainath · Karen Livescu

Fri Dec 11 06:50 AM -- 04:25 PM (PST) @ None
Event URL: https://neurips-sas-2020.github.io/ »

There is a trend in the machine learning community to adopt self-supervised approaches to pre-train deep networks. Self-supervised learning utilizes proxy supervised learning tasks, for example, distinguishing parts of the input signal from distractors, or generating masked input segments conditioned on the unmasked ones, to obtain training data from unlabeled corpora. These approaches make it possible to use a tremendous amount of unlabeled data on the web to train large networks and solve complicated tasks. ELMo, BERT, and GPT in NLP are famous examples in this direction. Recently self-supervised approaches for speech and audio processing are also gaining attention. These approaches combine methods for utilizing no or partial labels, unpaired text and audio data, contextual text and video supervision, and signals from user interactions. Although the research direction of self-supervised learning is active in speech and audio processing, current works are limited to several problems such as automatic speech recognition, speaker identification, and speech translation, partially due to the diversity of modeling in various speech and audio processing problems. There is still much unexplored territory in the research direction for self-supervised learning.

This workshop will bring concentrated discussions on self-supervision for the field of speech and audio processing via several invited talks, oral and poster sessions with high-quality papers, and a panel of leading researchers from academia and industry. Alongside research work on new self-supervised methods, data, applications, and results, this workshop will call for novel work on understanding, analyzing, and comparing different self-supervision approaches for speech and audio processing. The workshop aims to:
- Review existing and inspire new self-supervised methods and results,
- Motivate the application of self-supervision approaches to more speech and audio processing problems in academia and industry, and encourage discussion amongst experts and practitioners from the two realms,
- Encourage works on studying methods for understanding learned representations, comparing different self-supervision methods and comparing self-supervision to other self-training as well as transfer learning methods that low-resource speech and audio processing have long utilized,
- Facilitate communication within the field of speech and audio processing (e.g., people who attend conferences such as INTERSPEECH and ICASSP) as well as between the field and the whole machine learning community for sharing knowledge, ideas, and data, and encourage future collaboration to inspire innovation in the field and the whole community.

Fri 6:50 a.m. - 7:00 a.m.
Opening remarks (Introduction)
Hung-yi Lee
Fri 7:00 a.m. - 7:35 a.m.
Invited talk - A Broad Perspective into Self Supervised Learning for Speech Recognition (Invited talk)
Bhuvana Ramabhadran
Fri 7:35 a.m. - 7:45 a.m.
Q&A for invited talk - 1 (Q&A)
Fri 7:45 a.m. - 8:20 a.m.
Invited talk - Multimodal Distant Supervision (Invited talk) Video
Mark Hasegawa-Johnson
Fri 8:20 a.m. - 8:30 a.m.
Q&A for invited talk - Multimodal Distant Supervision (Q&A)
Fri 8:30 a.m. - 8:40 a.m.
Self-Supervised Learning using Contrastive Mixtures for Personalized Speech Enhancement (Contributed talk) Video
Aswin Sivaraman
Fri 8:40 a.m. - 8:50 a.m.
Self-supervised Pre-training Reduces Label Permutation Instability of Speech Separation (Contributed talk) Video
Sung-Feng Huang
Fri 8:50 a.m. - 9:00 a.m.
Augmentation adversarial training for self-supervised speaker recognition (Contributed talk) Video
jaesung Huh
Fri 9:00 a.m. - 9:10 a.m.
Neural Composition: Learning to Generate from Multiple Models (Contributed talk) Video
Denis Filimonov
Fri 9:10 a.m. - 9:20 a.m.
Towards Semi-Supervised Semantics Understanding from Speech (Contributed talk) Video
Cheng-I Lai
Fri 9:20 a.m. - 9:30 a.m.
The Zero Resource Speech Benchmark 2021. Metrics and baselines for unsupervised spoken language modeling (Contributed talk) Video
Tu Anh Nguyen
Fri 9:30 a.m. - 9:45 a.m.
Q&A for contributed talks between 11:30 and 12:30 (Q&A)
Fri 9:45 a.m. - 10:00 a.m.
Fri 10:00 a.m. - 10:35 a.m.
Invited talk - Speech Processing with Weak Supervision (Invited talk) Video
Dong Yu
Fri 10:35 a.m. - 10:45 a.m.
Q&A for invited talk - Speech Processing with Weak Supervision (Q&A)
Fri 10:45 a.m. - 10:55 a.m.
Towards Localisation of Keywords in Speech Using Weak Supervision (Contributed talk) Video
Kayode Olaleye
Fri 10:55 a.m. - 11:05 a.m.
Text-Free Image-to-Speech Synthesis Using Learned Segmental Units (Contributed talk) Video
Wei-Ning Hsu
Fri 11:05 a.m. - 11:15 a.m.
Self-Supervised Audio-Visual Separation of On-Screen Sounds from Unlabeled Videos (Contributed talk) Video
Efthymios Tzinis
Fri 11:15 a.m. - 11:25 a.m.
Multi-Format Contrastive Learning of Audio Representations (Contributed talk) Video
Aaron van den Oord
Fri 11:25 a.m. - 11:40 a.m.
Q&A for contributed talks between 1:45 and 2:25 (Q&A)
Fri 11:40 a.m. - 11:55 a.m.
Fri 11:55 a.m. - 12:30 p.m.
Invited talk - Underfitting and Uncertainty in Self-Supervised Predictive Models (Invited talk)
Chelsea Finn
Fri 12:30 p.m. - 12:40 p.m.
Q&A for invited talk - Underfitting and Uncertainty in Self-Supervised Predictive Models (Q&A)
Fri 12:40 p.m. - 1:15 p.m.
Invited talk - Towards robust self-supervised learning of speech representations (Invited talk) Video
Mirco Ravanelli
Fri 1:15 p.m. - 1:25 p.m.
Q&A for invited talk - Towards robust self-supervised learning of speech representations (Q&A)
Fri 1:25 p.m. - 1:35 p.m.
Similarity Analysis of Self-Supervised Speech Representations (Contributed talk) Video
Yu-An Chung
Fri 1:35 p.m. - 1:45 p.m.
Representation Learning for Sequence Data with Deep Autoencoding Predictive (Contributed talk) Video
Junwen Bai
Fri 1:45 p.m. - 1:55 p.m.
Pushing the Limits of Semi-Supervised Learning for Automatic Speech Recognition (Contributed talk) Video
Yu Zhang
Fri 1:55 p.m. - 2:05 p.m.
A Correspondence Variational Autoencoder for Unsupervised Acoustic Word Embedding (Contributed talk) Video
Puyuan Peng
Fri 2:05 p.m. - 2:15 p.m.
HUBERT: How much can a bad teacher benefit ASR pre-training? (Contributed talk) Video
Wei-Ning Hsu
Fri 2:15 p.m. - 2:30 p.m.
Q&A for contributed talks between 4:25 and 5:15 (Q&A)
Fri 2:30 p.m. - 2:45 p.m.
Fri 2:45 p.m. - 3:20 p.m.
Invited talk - Flexible contextualized speech representation learning for diverse downstream tasks (Invited talk) Video
Katrin Kirchhhoff
Fri 3:20 p.m. - 3:30 p.m.
Q&A for invited talk - Flexible contextualized speech representation learning for diverse downstream tasks (Q&A)
Fri 3:30 p.m. - 4:05 p.m.

De-noising auto-encoders can be pre-trained at a very large scale by noising and then reconstructing any input text. Existing methods, based on variations of masked languages models, have transformed the field and now provide the de facto initialization to be tuned for nearly every task. In this talk, I will present our work on sequence-to-sequence pre-training that introduces and carefully measures the impact of two new types of noising strategies. I will fist describe an approach that allows arbitrary noising, by learning to translate any corrupted text back to the original with standard Transformer-based neural machine translation architectures. I will show that the resulting mono-lingual (BART) and multi-lingual (mBART) models provide effective initialization for learning a wide range of discrimination and generation tasks, including question answer, summarization, and machine translation. I will also present our recently introduced MARGE model, where we self-supervise the reconstruction of target text by retrieving a set of related texts (in many languages) and conditioning on them to maximize the likelihood of generating the original. The objective noisily captures aspects of paraphrase, translation, multi-document summarization, and information retrieval, allowing for strong zero-shot performance with no fine-tuning, as well as consistent performance gain when fine tuned for individual tasks. Together, these techniques provide the most comprehensive set of pre-training methods to date, as well as the first viable alternative to the dominant masked language modeling pre-training paradigm.

Luke Zettlemoyer
Fri 4:05 p.m. - 4:15 p.m.
Q&A for invited talk - De-noising Sequence-to-Sequence Pre-training (Q&A)
Fri 4:15 p.m. - 4:25 p.m.
Closing remark (Introduction)
Abdelrahman Mohamed

Author Information

Abdelrahman Mohamed (Facebook AI Research (FAIR))
Hung-yi Lee (National Taiwan University)
Shinji Watanabe (Johns Hopkins University)
Shang-Wen Li (Amazon)
Tara Sainath (Google)
Karen Livescu (TTI-Chicago)

More from the Same Authors