A unified deep neural network for speaker and language recognition

By Fred Richardson, Douglas A. Reynolds, Najim Dehak. 2015

163 citations recorded by the source corpus.

Learned feature representations and sub-phoneme posteriors from Deep Neural Networks (DNNs) have been used separately to produce significant performance gains for speaker and language recognition tasks. In this work we show how these gains are possible using a single DNN for both speaker and language recognition. The unified DNN approach is shown to yield substantial performance improvements on the the 2013 Domain Adaptation Challenge speaker recognition task (55% reduction in EER for the out-of-domain condition) and on the NIST 2011 Language Recognition Evaluation (48% reduction in EER for the 30s test condition).

Subfields: Artificial Intelligence, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Jinyu Li
2. Dong Yu
3. Li Deng
4. Yoshua Bengio
5. Zhuo Chen
6. Mirco Ravanelli
7. Shinji Watanabe
8. Yifan Gong

In Signal Processing

1. Jinyu Li
2. Dong Yu
3. Yoshua Bengio
4. Mirco Ravanelli
5. Zhuo Chen
6. Yifan Gong
7. Yong Zhao
8. Frantisek Grézl

Citations in the corpus

• Exploring the Encoding Layer and Loss Function in End-to-End Speaker and Language Recognition System (2018, background): “• Local feature descriptors, which manifest as a variablelength feature sequence, include hand-crafted acoustic level features, such as log mel-filterbank energies (Fbank), mel-frequency cepstral coefficients (MFCC), perceptual linear prediction (PLP), shifted delta coefficients (SDC) features [2, 5”
• Interpretable Convolutional Filters with SincNet (2018, background): “employed in many state-of-the-art systems [19, 20, 21], directly feeding a CNN with spectrogram bins [22, 23, 24] or even with raw audio samples [25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37] is an approach of increasing popularity.”
• LID-Senones and Their Statistics for Language Identification (2018, methodology): “The baseline LID system used for comparison in this paper is the DNN/i-vector method [5], [9], [10] using statistics from bottleneck [7] features to form an i-vector [6], [11].”
• Multilingually trained bottleneck features in spoken language recognition (2017, methodology): “Combination of the i-vector paradigm with bottleneck features is a state-of-the-art technique for SLR (Ferrer et al., 2016; Jiang et al., 2014b; Richardson et al., 2015).”
• Speaker Recognition from Raw Waveform with SincNet (2018, methodology): “Most of past attempts, however, employed hand-crafted features such as FBANK and MFCC coefficients [12, 16, 17].”
• Speaker recognition based on deep learning: An overview (2020, methodology): “One can find that no matter whether we apply BNF alone [41] or concatenate it with other acoustic features [39], DNN-BNF/i-vector can significantly outperform the conventional GMM/i-vector, which indicates the effectiveness of the framework [36].”
• End-to-End Language Identification Using Attention-Based Recurrent Neural Networks (2016, background): “The deep bottleneck feed forward neural networks (DBN) that works as a front-end feature extractor has greatly enhanced the gaussian mixture modeluniversal background model (GMM-UBM) based i-Vector LID system [5, 6, 7, 8, 9, 10].”
• What Does the Speaker Embedding Encode? (2017, unknown): “Alternative approaches emerged, including DNN-based bottleneck feature extraction [10, 11, 12, 13] and direct speaker representation learning [14, 15, 16, 12].”
• End-to-End attention based text-dependent speaker verification (2016, methodology): “The baseline GMM-UBM and i-Vector systems are using the same MFCCs as KWS-DNNs.”