Short Utterance Compensation in Speaker Verification via Cosine-Based Teacher-Student Learning of Speaker Embeddings

Abstract

The short duration of an input utterance is one of the most critical threats that degrade the performance of speaker verification systems. This study aimed to develop an integrated text-independent speaker verification system that inputs utterances with short duration of 2 seconds or less. We propose an approach using a teacher-student learning framework for this goal, applied to short utterance compensation for the first time in our knowledge. The core concept of the proposed system is to conduct the compensation throughout the network that extracts the speaker embedding, mainly in phonetic-level, rather than compensating via a separate system after extracting the speaker embedding. In the proposed architecture, phoneticlevel features where each feature represents a segment of 130 ms are extracted using convolutional layers. A layer of gated recurrent units extracts an utterance-level feature using phoneticlevel features. The proposed approach also adopts a new objective function for teacher-student learning that considers both Kullback-Leibler divergence of output layers and cosine distance of speaker embeddings layers. Experiments were conducted using deep neural networks that take raw waveforms as input, and output speaker embeddings on VoxCeleb1 dataset. The proposed model could compensate approximately 65 % of the performance degradation due to the shortened duration.

Acoustic Scene Classification Using Teacher-Student Learning with Soft-Labels

“…The figure shows that the embeddings extracted from long segments have higher discriminative power. This superiority is also interpreted as that of short utterance compensation in the field of speaker verification [3,15].…”

“…With the inspiration from the previous researches [2,3], we explored two techniques to modify the TS learning to better conduct the ASC task. The first is extraction of soft-labels from multiple input segments.…”

“…This method enables extracting more general soft-labels and also includes the effect of data augmentation. The second is direct comparison of embeddings rather than the output layer, proposed in [3], was applied for the ASC task. We verified that the combination of the two techniques can further improve the performance of the ASC system.…”

Replay Attack Detection with Complementary High-Resolution Information Using End-to-End DNN for the ASVspoof 2019 Challenge

“…The DNN used in this study comprises convolutional neural networks (CNNs), gated recurrent units (GRUs) and fully connected layers (CNN-GRU) as used in [15][16][17]. In this architecture, input features are first processed using convolutional layers to extract frame-level embeddings.…”

“…A slightly modified ResNet was used for modeling the spectrograms, accounting for different stride sizes for time and frequency domains due to high-resolution in the frequency domain, and the number of residual blocks was adjusted to fit the provided ASV2019 physical access dataset. The raw waveform CNN-GRU model, proposed in [17], was used with a few modifications: one less residual block, a different specified input utterance length at training phase to fit the dataset, and additional loss functions for training (center loss [25] and speaker basis loss [26]). This model first extracts 128-dimensional frame-level representations using 1-dimensional convolutional layers.…”