Training Utterance-level Embedding Networks for Speaker Identification and Verification

Abstract

Encoding speaker-specific characteristics from speech signals into fixed length vectors is a key component of speaker identification and verification systems. This paper presents a deep neural network architecture for speaker embedding models where similarity in embedded utterance vectors explicitly approximates the similarity in vocal patterns of speakers. The proposed architecture contains an additional speaker embedding lookup table to compute loss based on embedding similarities. Furthermore, we propose a new feature sampling method for data augmentation. Experimentation based on two databases demonstrates that our model is more effective at speaker identification and verification when compared to a fully connected classifier and an end-to-end verification model.

A Deep Neural Network for Short-Segment Speaker Recognition

“…In [25], a long short-term memory (LSTM) architecture was applied on MFCC, resulting in an embedding used to verify the speaker of the utterance by means of cosine distance. Other attempts, such as the model proposed in [23] and [31], have used the LSTM architecture as an intermediate tool in extracting i-vectors. More sophisticated models combining CNN and RNN-based solutions were proposed in [26] and [24], applying several convolution layers in between the MFCC input and the RNN.…”

“…Recurrent neural networks (RNN) have been utilized in a number of studies. Recently, RNN models were employed in [23,31,25] with mel-frequency cepstrum coefficients (MFCC) as inputs. In [25], a long short-term memory (LSTM) architecture was applied on MFCC, resulting in an embedding used to verify the speaker of the utterance by means of cosine distance.…”

“…Emergence of challenging datasets such as Speaker In The Wild (SITW) [11] and its extended variations such as VoxCeleb1 and VoxCeleb2 [1,3], with more than 5,000 speakers and one million utterances, have enabled the opportunity to tackle speaker recognition in real-world scenarios. As a result, a number of deep learning solutions have been proposed for this purpose, including the models studied in [12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28].…”

H-VECTORS: Improving the robustness in utterance-level speaker embeddings using a hierarchical attention model

“…The goal of speaker recognition is to recognize a speaker from the characteristics of voices (Bai, Zhang, & Chen, 0000;Poddar, Sahidullah, & Saha, 2017). Representing the speaker properties into low dimensional feature space is beneficial for many downstream tasks, and such compact representations used to distinguish speakers (speaker embedding) have been an attractive topic and is widely used in some studies, such as speaker identification (Park, Cho, Park, Kim, & Park, 2018), verification (Le & Odobez, 2018;Novoselov, Shulipa, Kremnev, Kozlov, & Shchemelinin, 2018;Snyder, Garcia-Romero, Povey, & Khudanpur, 2017), detection (McLaren, Castan, Nandwana, Ferrer, & Yilmaz, 2018), segmentation (Garcia-Romero, Snyder, Sell, Povey, & McCree, 2017;Wang, Downey, Wan, Mansfield and Moreno, 2018), and speaker dependent speech enhancement (Chuang, Wang, Hung, Tsao, & Fang, 2019;Gao et al, 2015).…”

H-Vectors: Utterance-Level Speaker Embedding Using a Hierarchical Attention Model

“…The generation of compact representation used to distinguish speakers has been an attractive topic and widely used in some related studies, such as speaker identification [1], verification [2,3,4], detection [5], segmentation [6,7], and speaker dependent speech enhancement [8,9].…”