Abstract:Chunk-based inference stands out as a popular approach in developing real-time streaming speech recognition, valued for its simplicity and efficiency. However, because it restricts the model's focus to only the history and current chunk context, it may result in performance degradation in scenarios that demand consideration of future context. Addressing this, we propose a novel approach featuring Time-Shifted Contextual Attention (TSCA) and Dynamic Right Context (DRC) masking. Our method shows a relative word error rate reduction of 10 to 13.9% on the Librispeech dataset with the inclusion of in-context future information provided by TSCA. Moreover, we present a streaming automatic speech recognition pipeline that facilitates the integration of TSCA with minimal user-perceived latency, while also enabling batch processing capability, making it practical for various applications.
Abstract:RNN-Transducer (RNN-T) is a widely adopted architecture in speech recognition, integrating acoustic and language modeling in an end-to-end framework. However, the RNN-T predictor tends to over-rely on consecutive word dependencies in training data, leading to high deletion error rates, particularly with less common or out-of-domain phrases. Existing solutions, such as regularization and data augmentation, often compromise other aspects of performance. We propose SegAug, an alignment-based augmentation technique that generates contextually varied audio-text pairs with low sentence-level semantics. This method encourages the model to focus more on acoustic features while diversifying the learned textual patterns of its internal language model, thereby reducing deletion errors and enhancing overall performance. Evaluations on the LibriSpeech and Tedlium-v3 datasets demonstrate a relative WER reduction of up to 12.5% on small-scale and 6.9% on large-scale settings. Notably, most of the improvement stems from reduced deletion errors, with relative reductions of 45.4% and 18.5%, respectively. These results highlight SegAug's effectiveness in improving RNN-T's robustness, offering a promising solution for enhancing speech recognition performance across diverse and challenging scenarios.
Abstract:Deploying ASR models at an industrial scale poses significant challenges in hardware resource management, especially for long-form transcription tasks where audio may last for hours. Large Conformer models, despite their capabilities, are limited to processing only 15 minutes of audio on an 80GB GPU. Furthermore, variable input lengths worsen inefficiencies, as standard batching leads to excessive padding, increasing resource consumption and execution time. To address this, we introduce ChunkFormer, an efficient ASR model that uses chunk-wise processing with relative right context, enabling long audio transcriptions on low-memory GPUs. ChunkFormer handles up to 16 hours of audio on an 80GB GPU, 1.5x longer than the current state-of-the-art FastConformer, while also boosting long-form transcription performance with up to 7.7% absolute reduction on word error rate and maintaining accuracy on shorter tasks compared to Conformer. By eliminating the need for padding in standard batching, ChunkFormer's masked batching technique reduces execution time and memory usage by more than 3x in batch processing, substantially reducing costs for a wide range of ASR systems, particularly regarding GPU resources for models serving in real-world applications.
Abstract:In personalized Federated Learning (pFL), high data heterogeneity can cause significant gradient divergence across devices, adversely affecting the learning process. This divergence, especially when gradients from different users form an obtuse angle during aggregation, can negate progress, leading to severe weight and gradient update degradation. To address this issue, we introduce a new approach to pFL design, namely Federated Learning with Layer-wise Aggregation via Gradient Analysis (FedLAG), utilizing the concept of gradient conflict at the layer level. Specifically, when layer-wise gradients of different clients form acute angles, those gradients align in the same direction, enabling updates across different clients toward identifying client-invariant features. Conversely, when layer-wise gradient pairs make create obtuse angles, the layers tend to focus on client-specific tasks. In hindsights, FedLAG assigns layers for personalization based on the extent of layer-wise gradient conflicts. Specifically, layers with gradient conflicts are excluded from the global aggregation process. The theoretical evaluation demonstrates that when integrated into other pFL baselines, FedLAG enhances pFL performance by a certain margin. Therefore, our proposed method achieves superior convergence behavior compared with other baselines. Extensive experiments show that our FedLAG outperforms several state-of-the-art methods and can be easily incorporated with many existing methods to further enhance performance.