Recordings of team presentations at NeurIPS 2021 are available for viewing.
- AMAAI Lab
- Logitech AI
- MARL + Soundsensing
- Stellenbosch LSL
Team information will be updated as it becomes available.
Khaled Koutini1, Jan Schlüter1, Hamid Eghbal-zadeh1,2, Gerhard Widmer1,2
1Institute of Computational Perception, Johannes Kepler University Linz, Austria
2LIT AI Lab, Johannes Kepler University Linz, Austria
The great success of transformer-based models in natural language processing (NLP) has led to various attempts at adapting these architectures to other domains such as vision and audio. Recent work has shown that transformers can outperform Convolutional Neural Networks (CNNs) on vision and audio tasks. However, one of the main shortcomings of transformer models, compared to the well-established CNNs, is the computational complexity. Compute and memory complexity grow quadratically with the input length. Therefore, there has been extensive work on optimizing transformers, but often at the cost of lower predictive performance. In this work, we propose a novel method to optimize and regularize transformers on audio spectrograms. The proposed models achieve a new state-of-the-art performance on Audioset and can be trained on a single consumer-grade GPU. Furthermore, we propose a transformer model that outperforms CNNs in terms of both performance and training speed.
Ho-Hsiang Wu1, Prem Seetharaman2
1Music and Audio Research Laboratory (MARL) New York University
We propose Wav2CLIP, a robust audio representation learning method by distilling from Contrastive Language-Image Pre-training (CLIP). Wav2CLIP projects audio into a shared embedding space with images and text, which enables multimodal applications such as zero-shot classification, and cross-modal retrieval. Furthermore, Wav2CLIP needs just ~10% of the data to achieve competitive performance on downstream tasks compared with fully supervised models, and is more efficient to pre-train than competing methods as it does not require learning a visual model in concert with an auditory model.
Mashrur Mahmud Morshed1,2, Ahmad Omar Ahsan1
1AI Engineer, Intelligent Machines Ltd.
2Systems & Software Lab, IUT
Our submission involves using lightweight all-MLP architectures on spectrogram inputs, and generating scene embeddings by concatenation followed by temporal interpolation. We have two submission variants: kwmlp (Keyword-MLP) and audiomlp.
Keyword-MLP† is a model for keyword spotting on Google Speech Commands V2-35. It consists of identical, sequentially stacked gated MLP (gMLP‡) blocks and accepts MFCCs as inputs. The model has an isotropic architecture: all blocks are identical and have the same input and output size of (T, 64). The timestamp embeddings of size 64 are obtained by simply removing the classification head at the end.
The audiomlp submission consists of the Audio-MLP-Autoencoder, a denoising autoencoder where the encoder and decoder both are similar sequentially stacked gMLP blocks. The ‘noise’ for the autoencoder is Spectral Augmentation, i.e. zeroed out time and frequency bands. There is a narrow bottleneck between the encoder and the decoder, which outputs the timestamp embeddings of size 8.
Both approaches use the same scene embedding algorithm. The raw audios of arbitrary duration are split into 1s chunks and then converted to spectrograms (with a 10ms hop length) and passed to the models. The output timestamp embeddings for each 1s chunk are first concatenated across the time axis. We then reduce the time dimension with linear interpolation and finally flatten to obtain the scene embedding vector. As opposed to the naive averaging approach, we tried interpolation with the aim of conserving temporal information in the scene embeddings.
 Morshed, Mashrur M., and Ahmad Omar Ahsan. "Attention-Free Keyword Spotting." arXiv preprint arXiv:2110.07749 (2021). †Under review at ICASSP-2022.
 Liu, Hanxiao, Zihang Dai, David R. So, and Quoc V. Le. "Pay Attention to MLPs." arXiv preprint arXiv:2105.08050 (2021). ‡Accepted to NeurIPS-2021.
Tung-Yu Wu (EE)1, Chen-An Li (CSIE)1, Tzu-Han Lin (CSIE)1, Tsu-Yuan Hsu (CSIE)1, advised by Hung-Yi Lee1
1National Taiwan University
Rather than focusing on the development of one single huge SSL models, we investigate the collaboration of multiple models for feature representation. Experiments are conducted to examine the effects of intra- and inter-model feature operations such as early fusion and concatenation.
1Xiaomi AI Lab.
Our system consists of an EfficientNet trained on the publicly available Audioset dataset using weak supervision. Our approach focuses on using an EfficientNet-B2 with decision-level aggregation to embed an input audio sequence into a fixed sized 320ms long embedding. Different from other contestants on which trained their models on Audioset, our proposed approach is lightweight in nature (8 Million parameters), while achieving a similar or better performance in most sound-based evaluation tasks (i.e., ESC-50, FSDk50, CREMA-D).