GE2E-AC: Generalized End-to-End Loss Training for Accent Classification

TL;DR

A GE2E-AC is proposed, in which a model is trained to extract accent embedding or AE of an input utterance such that the AEs of the same accent class get closer, instead of directly minimizing the classification loss.

Abstract

Accent classification or AC is a task to predict the accent type of an input utterance, and it can be used as a preliminary step toward accented speech recognition and accent conversion. Existing studies have often achieved such classification by training a neural network model to minimize the classification error of the predicted accent label, which can be obtained as a model output. Since we optimize the entire model only from the perspective of classification loss during training time in this approach, the model might learn to predict the accent type from irrelevant features, such as individual speaker identity, which are not informative during test time. To address this problem, we propose a GE2E-AC, in which we train a model to extract accent embedding or AE of an input utterance such that the AEs of the same accent class get closer, instead of directly minimizing the classification loss. We experimentally show the effectiveness of the proposed GE2E-AC, compared to the baseline model trained with the conventional cross-entropy-based loss.

Figures (2)

• Figure 1: The network architectures of (a) the baseline CE-based and (b) the proposed GE2E-based accent classification. For arbitrary vector $\bm{\phi} \in \mathbb{R}^K$, $\mathrm{Context}(\bm{\phi})$ indicates a function $f_{\bm{\phi}}: \mathbb{R}^{T \times F'} \mapsto \mathbb{R}^{T \times KF'}$, which maps matrix $X \in \mathbb{R}^{T \times F'}$ to $f_{\bm{\phi}}(X) = \left[ M^{(1)} X, \cdots, M^{(K)} X \right]$, where $M^{(k)} \in \mathbb{R}^{T \times T}$ is a matrix each of whose entries $M^{(k)}_{ij}$ is one if $j = i + \phi_k$ and zero otherwise for all $k \in \{ 1, \dots, K \}$. $\mathrm{CatMeanStd}$ is a function $f: \mathbb{R}^{T \times F'} \mapsto \mathbb{R}^{2F'}$, which maps matrix $X \in \mathbb{R}^{T \times F'}$ to $\left[\bm{\mu}^\mathsf{T}, \bm{\sigma}^\mathsf{T}\right]^\mathsf{T}$, where $\bm{\mu} \in \mathbb{R}^{F'}$ and $\bm{\sigma} \in \mathbb{R}^{F'}$ are the vectors each of whose entries $\mu_k$ and $\sigma_k$ are the mean and the standard deviation of vector $\left[ X_{1k}, \dots, X_{Tk} \right]^\mathsf{T}$, respectively.
• Figure 2: Confusion matrices with three accent types. The value plotted in each entry indicates the number of corresponding samples. "En," "Am," and "Sc" stand for English, American, and Scottish, respectively.