Sarthak Yadav

Sarthak Yadav, Sergios Theodoridis, Lars Kai Hansen et al.

In this work, we propose a Multi-Window Masked Autoencoder (MW-MAE) fitted with a novel Multi-Window Multi-Head Attention (MW-MHA) module that facilitates the modelling of local-global interactions in every decoder transformer block through attention heads of several distinct local and global windows. Empirical results on ten downstream audio tasks show that MW-MAEs consistently outperform standard MAEs in overall performance and learn better general-purpose audio representations, along with demonstrating considerably better scaling characteristics. Investigating attention distances and entropies reveals that MW-MAE encoders learn heads with broader local and global attention. Analyzing attention head feature representations through Projection Weighted Canonical Correlation Analysis (PWCCA) shows that attention heads with the same window sizes across the decoder layers of the MW-MAE learn correlated feature representations which enables each block to independently capture local and global information, leading to a decoupled decoder feature hierarchy. Code for feature extraction and downstream experiments along with pre-trained models will be released publically.

Sarthak Yadav, Neil Zeghidour

Deep audio classification, traditionally cast as training a deep neural network on top of mel-filterbanks in a supervised fashion, has recently benefited from two independent lines of work. The first one explores "learnable frontends", i.e., neural modules that produce a learnable time-frequency representation, to overcome limitations of fixed features. The second one uses self-supervised learning to leverage unprecedented scales of pre-training data. In this work, we study the feasibility of combining both approaches, i.e., pre-training learnable frontend jointly with the main architecture for downstream classification. First, we show that pretraining two previously proposed frontends (SincNet and LEAF) on Audioset drastically improves linear-probe performance over fixed mel-filterbanks, suggesting that learnable time-frequency representations can benefit self-supervised pre-training even more than supervised training. Surprisingly, randomly initialized learnable filterbanks outperform mel-scaled initialization in the self-supervised setting, a counter-intuitive result that questions the appropriateness of strong priors when designing learnable filters. Through exploratory analysis of the learned frontend components, we uncover crucial differences in properties of these frontends when used in a supervised and self-supervised setting, especially the affinity of self-supervised filters to diverge significantly from the mel-scale to model a broader range of frequencies.

Sarthak Yadav, Sergios Theodoridis, Zheng-Hua Tan

In recent years, self-supervised learning has amassed significant interest for training deep neural representations without labeled data. One such self-supervised learning approach is masked spectrogram modeling, where the objective is to learn semantically rich contextual representations by predicting removed or hidden portions of the input audio spectrogram. With the Transformer neural architecture at its core, masked spectrogram modeling has emerged as the prominent approach for learning general purpose audio representations, a.k.a. audio foundation models. Meanwhile, addressing the issues of the Transformer architecture, in particular the underlying Scaled Dot-product Attention operation, which scales quadratically with input sequence length, has led to renewed interest in recurrent sequence modeling approaches. Among them, Selective structured state space models (such as Mamba) and extended Long Short-Term Memory (xLSTM) are the two most promising approaches which have experienced widespread adoption. While the body of work on these two topics continues to grow, there is currently a lack of an adequate overview encompassing the intersection of these topics. In this paper, we present a comprehensive overview of the aforementioned research domains, covering masked spectrogram modeling and the previously mentioned neural sequence modeling architectures, Mamba and xLSTM. Further, we compare Transformers, Mamba and xLSTM based masked spectrogram models in a unified, reproducible framework on ten diverse downstream audio classification tasks, which will help interested readers to make informed decisions regarding suitability of the evaluated approaches to adjacent applications.

Sarthak Yadav, Sergios Theodoridis, Zheng-Hua Tan

Masked Autoencoders (MAEs) trained on audio spectrogram patches have emerged as a prominent approach for learning self-supervised audio representations. While several recent papers have evaluated key aspects of training MAEs on audio data, the majority of these approaches still leverage vanilla transformer building blocks, whereas the transformer community has seen steady integration of newer architectural advancements. In this work, we propose AudioMAE++, a revamped audio masked autoencoder with two such enhancements, namely macaron-style transformer blocks with gated linear units. When pretrained on the AudioSet dataset, the proposed AudioMAE++ models outperform existing MAE based approaches on 10 diverse downstream tasks, demonstrating excellent performance on audio classification and speech-based benchmarks. The proposed AudioMAE++ models also demonstrate excellent scaling characteristics, outperforming directly comparable standard MAE baselines with up to 4x more parameters.

Sarthak Yadav, Zheng-Hua Tan

Despite its widespread adoption as the prominent neural architecture, the Transformer has spurred several independent lines of work to address its limitations. One such approach is selective state space models, which have demonstrated promising results for language modelling. However, their feasibility for learning self-supervised, general-purpose audio representations is yet to be investigated. This work proposes Audio Mamba, a selective state space model for learning general-purpose audio representations from randomly masked spectrogram patches through self-supervision. Empirical results on ten diverse audio recognition downstream tasks show that the proposed models, pretrained on the AudioSet dataset, consistently outperform comparable self-supervised audio spectrogram transformer (SSAST) baselines by a considerable margin and demonstrate better performance in dataset size, sequence length and model size comparisons.

Sarthak Yadav, Mary Ellen Foster

Most of the existing isolated sound event datasets comprise a small number of sound event classes, usually 10 to 15, restricted to a small domain, such as domestic and urban sound events. In this work, we introduce GISE-51, a dataset spanning 51 isolated sound events belonging to a broad domain of event types. We also release GISE-51-Mixtures, a dataset of 5-second soundscapes with hard-labelled event boundaries synthesized from GISE-51 isolated sound events. We conduct baseline sound event recognition (SER) experiments on the GISE-51-Mixtures dataset, benchmarking prominent convolutional neural networks, and models trained with the dataset demonstrate strong transfer learning performance on existing audio recognition benchmarks. Together, GISE-51 and GISE-51-Mixtures attempt to address some of the shortcomings of recent sound event datasets, providing an open, reproducible benchmark for future research along with the freedom to adapt the included isolated sound events for domain-specific applications.

Sayan Mandal, Sarthak Yadav, Atul Rai

Bengali is a prominent language of the Indian subcontinent. However, while many state-of-the-art acoustic models exist for prominent languages spoken in the region, research and resources for Bengali are few and far between. In this work, we apply CTC based CNN-RNN networks, a prominent deep learning based end-to-end automatic speech recognition technique, to the Bengali ASR task. We also propose and evaluate the applicability and efficacy of small 7x3 and 3x3 convolution kernels which are prominently used in the computer vision domain primarily because of their FLOPs and parameter efficient nature. We propose two CNN blocks, 2-layer Block A and 4-layer Block B, with the first layer comprising of 7x3 kernel and the subsequent layers comprising solely of 3x3 kernels. Using the publicly available Large Bengali ASR Training data set, we benchmark and evaluate the performance of seven deep neural network configurations of varying complexities and depth on the Bengali ASR task. Our best model, with Block B, has a WER of 13.67, having an absolute reduction of 1.39% over comparable model with larger convolution kernels of size 41x11 and 21x11.

Sarthak Yadav, Atul Rai

Majority of the recent approaches for text-independent speaker recognition apply attention or similar techniques for aggregation of frame-level feature descriptors generated by a deep neural network (DNN) front-end. In this paper, we propose methods of convolutional attention for independently modelling temporal and frequency information in a convolutional neural network (CNN) based front-end. Our system utilizes convolutional block attention modules (CBAMs) [1] appropriately modified to accommodate spectrogram inputs. The proposed CNN front-end fitted with the proposed convolutional attention modules outperform the no-attention and spatial-CBAM baselines by a significant margin on the VoxCeleb [2, 3] speaker verification benchmark, and our best model achieves an equal error rate of 2:031% on the VoxCeleb1 test set, improving the existing state of the art result by a significant margin. For a more thorough assessment of the effects of frequency and temporal attention in real-world conditions, we conduct ablation experiments by randomly dropping frequency bins and temporal frames from the input spectrograms, concluding that instead of modelling either of the entities, simultaneously modelling temporal and frequency attention translates to better real-world performance.