Gaussian Kernelized Self-Attention for Long Sequence Data and its Application to CTC-Based Speech Recognition

Abstract

Self-attention (SA) based models have recently achieved significant performance improvements in hybrid and end-to-end automatic speech recognition (ASR) systems owing to their flexible context modeling capability. However, it is also known that the accuracy degrades when applying SA to long sequence data. This is mainly due to the length mismatch between the inference and training data because the training data are usually divided into short segments for efficient training. To mitigate this mismatch, we propose a new architecture, which is a variant of the Gaussian kernel, which itself is a shift-invariant kernel. First, we mathematically demonstrate that self-attention with shared weight parameters for queries and keys is equivalent to a normalized kernel function. By replacing this kernel function with the proposed Gaussian kernel, the architecture becomes completely shift-invariant with the relative position information embedded using a frame indexing technique. The proposed Gaussian kernelized SA was applied to connectionist temporal classification (CTC) based ASR. An experimental evaluation with the Corpus of Spontaneous Japanese (CSJ) and TEDLIUM 3 benchmarks shows that the proposed SA achieves a significant improvement in accuracy (e.g., from 24.0% WER to 6.0% in CSJ) in long sequence data without any windowing techniques.

QSAN: A Near-term Achievable Quantum Self-Attention Network

“…Although SAM is very beneficial, its complexity escalates quadratically with the length of the input sequence, severely hindering its exploration on longer sequences. Some proposals, such as Nyström matrix decomposition [11], kernel methods [12,13], hashing strategies [14], sparsification [15][16][17], random projections [18], etc., have good effect in diminishing the complexity. In further, Ref.…”

Document Layout Analysis Via Positional Encoding

“…Positional encoding is an important part of the Transformer [14] and the bottom input component of the encoder and decoder stack. Positional encoding has been applied in many tasks and has played an irreplaceable role [15,16,17,18,19,20,21]. At present, the task of document layout analysis pays little attention to solving the continuity of the classification document prediction area and maintaining the detailed information of the image, especially the use of positional encoding for document layout analysis.…”