Federated learning (FL) allows clients to collaboratively train a global model without sharing their local data with a server. However, clients' contributions to the server can still leak sensitive information. Differential privacy (DP) addresses such leakage by providing formal privacy guarantees, with mechanisms that add randomness to the clients' contributions. The randomness makes it infeasible to train large transformer-based models, common in modern federated learning systems. In this work, we empirically evaluate the practicality of fine-tuning large scale on-device transformer-based models with differential privacy in a federated learning system. We conduct comprehensive experiments on various system properties for tasks spanning a multitude of domains: speech recognition, computer vision (CV) and natural language understanding (NLU). Our results show that full fine-tuning under differentially private federated learning (DP-FL) generally leads to huge performance degradation which can be alleviated by reducing the dimensionality of contributions through parameter-efficient fine-tuning (PEFT). Our benchmarks of existing DP-PEFT methods show that DP-Low-Rank Adaptation (DP-LoRA) consistently outperforms other methods. An even more promising approach, DyLoRA, which makes the low rank variable, when naively combined with FL would straightforwardly break differential privacy. We therefore propose an adaptation method that can be combined with differential privacy and call it DP-DyLoRA. Finally, we are able to reduce the accuracy degradation and word error rate (WER) increase due to DP to less than 2% and 7% respectively with 1 million clients and a stringent privacy budget of $ε=2$.

Automatic Speech Recognition (ASR) is widely used within consumer devices such as mobile phones. Recently, personalization or on-device model fine-tuning has shown that adaptation of ASR models towards target user speech improves their performance over rare words or accented speech. Despite these gains, fine-tuning on user data (target domain) risks the personalized model to forget knowledge about its original training distribution (source domain) i.e. catastrophic forgetting, leading to subpar general ASR performance. A simple and efficient approach to combat catastrophic forgetting is to measure forgetting via a validation set that represents the source domain distribution. However, such validation sets are large and impractical for mobile devices. Towards this, we propose a novel method to subsample a substantially large validation set into a smaller one while maintaining the ability to estimate forgetting. We demonstrate the efficacy of such a dataset in mitigating forgetting by utilizing it to dynamically determine the number of ideal fine-tuning epochs. When measuring the deviations in per user fine-tuning epochs against a 50x larger validation set (oracle), our method achieves a lower mean-absolute-error (3.39) compared to randomly selected subsets of the same size (3.78-8.65). Unlike random baselines, our method consistently tracks the oracle’s behaviour across three different forgetting thresholds.

Data augmentation (DA) is ubiquitously used in training of Automatic Speech Recognition (ASR) models. DA offers increased data variability, robustness and generalization against different acoustic distortions. Recently, personalization of ASR models on mobile devices has been shown to improve Word Error Rate (WER). This paper evaluates data augmentation in this context and proposes persoDA; a DA method driven by user’s data utilized to personalize ASR [1] –[3]. persoDA aims to augment training with data specifically tuned towards acoustic characteristics of the end-user, as opposed to standard augmentation based on Multi-Condition Training (MCT) that applies random reverberation and noises. Our evaluation with an ASR conformer-based baseline trained on Librispeech and per-sonalized for VOICES [4] shows that persoDA achieves a 13.9% relative WER reduction over using standard data augmentation (using random noise & reverberation). Furthermore, persoDA shows 16% to 20% faster convergence over MCT.

Diffusion based Text-To-Music (TTM) models generate music corresponding to text descriptions. Typically UNet based diffusion models condition on text embeddings generated from a pre-trained large language model or from a cross-modality audio-language representation model. This work proposes a diffusion based TTM, in which the UNet is conditioned on both (i) a uni-modal language model (e.g., T5) via cross-attention and (ii) a cross-modal audio-language representation model (e.g., CLAP) via Feature-wise Linear Modulation (FiLM). The diffusion model is trained to exploit both a local text representation from the T5 and a global representation from the CLAP. Furthermore, we propose modifications that extract both global and local representations from the T5 through pooling mechanisms that we call mean pooling and self-attention pooling. This approach mitigates the need for an additional encoder (e.g., CLAP) to extract a global representation, thereby reducing the number of model parameters. Our results show that incorporating the CLAP global embeddings to the T5 local embeddings enhances text adherence (KL=1.47) compared to a baseline model solely relying on the T5 local embeddings (KL=1.54). Alternatively, extracting global text embeddings directly from the T5 local embeddings through the proposed mean pooling approach yields superior generation quality (FAD=1.89) while exhibiting marginally inferior text adherence (KL=1.51) against the model conditioned on both CLAP and T5 text embeddings (FAD=1.94 and KL=1.47). Our proposed solution is not only efficient but also compact in terms of the number of parameters required.

Automatic Speech Recognition (ASR) still face challenges when recognizing time-variant rare-phrases. Contextual biasing (CB) modules bias ASR model towards such contextually-relevant phrases. During training, a list of biasing phrases are selected from a large pool of phrases following a sampling strategy. In this work we firstly analyse different sampling strategies to provide insights into the training of CB for ASR with correlation plots between the bias embeddings among various training stages. Secondly, we introduce a neighbourhood attention (NA) that localizes self attention (SA) to the nearest neighbouring frames to further refine the CB output. The results show that this proposed approach provides on average a 25.84% relative WER improvement on LibriSpeech sets and rare-word evaluation compared to the baseline.

On-device Automatic Speech Recognition (ASR) models trained on speech data of a large population might underperform for individuals unseen during training. This is due to a domain shift between user data and the original training data, differed by user’s speaking characteristics and environmental acoustic conditions. ASR personalisation is a solution that aims to exploit user data to improve model robustness. The majority of ASR personalisation methods assume labelled user data for supervision. Personalisation without any labelled data is challenging due to limited data size and poor quality of recorded audio samples. This work addresses unsupervised personalisation by developing a novel consistency based training method via pseudo-labelling. Our method achieves a relative Word Error Rate Reduction (WERR) of 17.3% on unlabelled training data and 8.1% on held-out data compared to a pre-trained model, and outperforms the current state-of-the art methods.

State-of-the-art edge devices are capable of not only inferring machine learning (ML) models but also training them on the device with local data. When this local data is sensitive, privacy becomes a crucial property that must be addressed. This implies that sharing data with a server for training a model is undesirable and should be avoided. The Federated Learning (FL) approach can help in these situations, however, FL alone is still not the ultimate tool to solve all challenges, especially when privacy is a major concern. We propose a privacy-preserving FL framework, which leverages the concepts of bitwise quantization, local differential privacy (LDP), and feature hashing for input representation in the collaborative training of ML models. In our approach, the local model updates are first quantized, then a randomized-response technique is applied on the resulting update vector. Although our proposed framework functions with arbitrary types of input features, we emphasize its usability with natural language data. The text input on the client-side is encoded using a rolling-hash-based representation, which provides a combined solution for the high resource demands of embedding algorithms and the privacy concerns of sharing sensitive data. We evaluate our method in a sentiment analysis task using the IMDB Movie Reviews dataset as well as a rating prediction task with the MovieLens dataset augmented with additional movie keywords. We demonstrate that our approach is a feasible solution for private language processing tasks on edge devices without the use of resource-hungry language models or privacy-violating collection of client data.

We propose a novel Patched Multi-Condition Training (pMCT) method for robust Automatic Speech Recognition (ASR).pMCT employs Multi-condition Audio Modification and Patching (MAMP) via mixing patches of the same utterance extracted from clean and distorted speech.Training using patchmodified signals improves robustness of models in noisy reverberant scenarios.Our proposed pMCT is evaluated on the LibriSpeech dataset showing improvement over using vanilla Multi-Condition Training (MCT).For analyses on robust ASR, we employed pMCT on the VOiCES dataset which is a noisy reverberant dataset created using utterances from LibriSpeech.In the analyses, pMCT achieves 23.1% relative WER reduction compared to the MCT.