Jan Silovský

Martin Pelikan, Sheikh Shams Azam, Vitaly Feldman et al. · apple-ml

While federated learning (FL) and differential privacy (DP) have been extensively studied, their application to automatic speech recognition (ASR) remains largely unexplored due to the challenges in training large transformer models. Specifically, large models further exacerbate issues in FL as they are particularly susceptible to gradient heterogeneity across layers, unlike the relatively uniform gradient behavior observed in shallow models. As a result, prior works struggle to converge with standard optimization techniques, even in the absence of DP mechanisms. To the best of our knowledge, no existing work establishes a competitive, practical recipe for FL with DP in the context of ASR. To address this gap, we establish \textbf{the first benchmark for FL with DP in end-to-end ASR}. Our approach centers on per-layer clipping and layer-wise gradient normalization: theoretical analysis reveals that these techniques together mitigate clipping bias and gradient heterogeneity across layers in deeper models. Consistent with these theoretical insights, our empirical results show that FL with DP is viable under strong privacy guarantees, provided a population of at least several million users. Specifically, we achieve user-level (7.2, $10^{-9}$)-DP (resp. (4.5, $10^{-9}$)-DP) with only a 1.3% (resp. 4.6%) absolute drop in word error rate when extrapolating to high (resp. low) population scales for FL with DP in ASR. Although our experiments focus on ASR, the underlying principles we uncover - particularly those concerning gradient heterogeneity and layer-wise gradient normalization - offer broader guidance for designing scalable, privacy-preserving FL algorithms for large models across domains. Code of all experiments and benchmarks is available at https://github.com/apple/ml-pfl4asr.

Sheikh Shams Azam, Tatiana Likhomanenko, Martin Pelikan et al. · apple-ml

In this paper, we start by training End-to-End Automatic Speech Recognition (ASR) models using Federated Learning (FL) and examining the fundamental considerations that can be pivotal in minimizing the performance gap in terms of word error rate between models trained using FL versus their centralized counterpart. Specifically, we study the effect of (i) adaptive optimizers, (ii) loss characteristics via altering Connectionist Temporal Classification (CTC) weight, (iii) model initialization through seed start, (iv) carrying over modeling setup from experiences in centralized training to FL, e.g., pre-layer or post-layer normalization, and (v) FL-specific hyperparameters, such as number of local epochs, client sampling size, and learning rate scheduler, specifically for ASR under heterogeneous data distribution. We shed light on how some optimizers work better than others via inducing smoothness. We also summarize the applicability of algorithms, trends, and propose best practices from prior works in FL (in general) toward End-to-End ASR models.

Man-Ling Sung, Jan Silovsky, Man-Hung Siu et al.

Transfer learning and knowledge distillation has recently gained a lot of attention in the deep learning community. One transfer approach, the student-teacher learning, has been shown to successfully create ``small'' student neural networks that mimic the performance of a much bigger and more complex ``teacher'' networks. In this paper, we investigate an extension to this approach and transfer from a non-neural-based machine learning pipeline as teacher to a neural network (NN) student, which would allow for joint optimization of the various pipeline components and a single unified inference engine for multiple ML tasks. In particular, we explore replacing the random forest classifier by transfer learning to a student NN. We experimented with various NN topologies on 100 OpenML tasks in which random forest has been one of the best solutions. Our results show that for the majority of the tasks, the student NN can indeed mimic the teacher if one can select the right NN hyper-parameters. We also investigated the use of random forest for selecting the right NN hyper-parameters.

Jan Silovsky, Liuhui Deng, Arturo Argueta et al.

Voice technology has become ubiquitous recently. However, the accuracy, and hence experience, in different languages varies significantly, which makes the technology not equally inclusive. The availability of data for different languages is one of the key factors affecting accuracy, especially in training of all-neural end-to-end automatic speech recognition systems. Cross-lingual knowledge transfer and iterative pseudo-labeling are two techniques that have been shown to be successful for improving the accuracy of ASR systems, in particular for low-resource languages, like Ukrainian. Our goal is to train an all-neural Transducer-based ASR system to replace a DNN-HMM hybrid system with no manually annotated training data. We show that the Transducer system trained using transcripts produced by the hybrid system achieves 18% reduction in terms of word error rate. However, using a combination of cross-lingual knowledge transfer from related languages and iterative pseudo-labeling, we are able to achieve 35% reduction of the error rate.

Zhuolin Jiang, Jan Silovsky, Man-Hung Siu et al.

Multi-label image classification has generated significant interest in recent years and the performance of such systems often suffers from the not so infrequent occurrence of incorrect or missing labels in the training data. In this paper, we extend the state-of the-art of training classifiers to jointly deal with both forms of errorful data. We accomplish this by modeling noisy and missing labels in multi-label images with a new Noise Modeling Network (NMN) that follows our convolutional neural network (CNN), integrates with it, forming an end-to-end deep learning system, which can jointly learn the noise distribution and CNN parameters. The NMN learns the distribution of noise patterns directly from the noisy data without the need for any clean training data. The NMN can model label noise that depends only on the true label or is also dependent on the image features. We show that the integrated NMN/CNN learning system consistently improves the classification performance, for different levels of label noise, on the MSR-COCO dataset and MSR-VTT dataset. We also show that noise performance improvements are obtained when multiple instance learning methods are used.

Andrew Titus, Jan Silovsky, Nanxin Chen et al.

Spoken language identification (LID) technologies have improved in recent years from discriminating largely distinct languages to discriminating highly similar languages or even dialects of the same language. One aspect that has been mostly neglected, however, is discrimination of languages for multilingual speakers, despite being a primary target audience of many systems that utilize LID technologies. As we show in this work, LID systems can have a high average accuracy for most combinations of languages while greatly underperforming for others when accented speech is present. We address this by using coarser-grained targets for the acoustic LID model and integrating its outputs with interaction context signals in a context-aware model to tailor the system to each user. This combined system achieves an average 97% accuracy across all language combinations while improving worst-case accuracy by over 60% relative to our baseline.

Herbert Gish, Jan Silovsky, Man-Ling Sung et al.

We investigate the problem of machine learning with mislabeled training data. We try to make the effects of mislabeled training better understood through analysis of the basic model and equations that characterize the problem. This includes results about the ability of the noisy model to make the same decisions as the clean model and the effects of noise on model performance. In addition to providing better insights we also are able to show that the Maximum Likelihood (ML) estimate of the parameters of the noisy model determine those of the clean model. This property is obtained through the use of the ML invariance property and leads to an approach to developing a classifier when training has been mislabeled: namely train the classifier on noisy data and adjust the decision threshold based on the noise levels and/or class priors. We show how our approach to mislabeled training works with multi-layered perceptrons (MLPs).

K. Youssef, Louis-S. Bouchard, K. Z. Haigh et al.

With the development and widespread use of wireless devices in recent years (mobile phones, Internet of Things, Wi-Fi), the electromagnetic spectrum has become extremely crowded. In order to counter security threats posed by rogue or unknown transmitters, it is important to identify RF transmitters not by the data content of the transmissions but based on the intrinsic physical characteristics of the transmitters. RF waveforms represent a particular challenge because of the extremely high data rates involved and the potentially large number of transmitters present in a given location. These factors outline the need for rapid fingerprinting and identification methods that go beyond the traditional hand-engineered approaches. In this study, we investigate the use of machine learning (ML) strategies to the classification and identification problems, and the use of wavelets to reduce the amount of data required. Four different ML strategies are evaluated: deep neural nets (DNN), convolutional neural nets (CNN), support vector machines (SVM), and multi-stage training (MST) using accelerated Levenberg-Marquardt (A-LM) updates. The A-LM MST method preconditioned by wavelets was by far the most accurate, achieving 100% classification accuracy of transmitters, as tested using data originating from 12 different transmitters. We discuss strategies for extension of MST to a much larger number of transmitters.