Publications

In 2019, the U.S. National Institute of Standards and Technology (NIST) conducted a leaderboard style speaker recognition challenge using conversational telephone speech (CTS) data extracted from the unexposed portion of the Call My Net 2 (CMN2) corpus previously used in the 2018 Speaker Recognition Evaluation (SRE). The SRE19 CTS Challenge was organized in a similar manner to SRE18, except it offered only the open training condition. In addition, similar to the NIST i-vector challenge, the evaluation set consisted of two subsets: a progress subset, and a test subset. The progress subset comprised 30% of the trials and was used to monitor progress on the leaderboad, while the remaining 70% of the trials formed the test subset, which was used to generate the official final results determined at the end of the challenge. Which subset (i.e., progress or test) a trial belonged to was unknown to challenge participants, and each system submission had to contain outputs for all of trials. The CTS Challenge also served as a prerequisite for entrance to the main SRE19 whose primary task was audio-visual person recognition. A total of 67 organizations (forming 51 teams) from academia and industry participated in the CTS Challenge and submitted 1347 valid system outputs. This paper presents an overview of the evaluation and several analyses of system performance for all primary conditions in the CTS Challenge. Compared to the CTS track of the SRE18, the SRE19 CTS Challenge results indicate remarkable improvements in performance which are mainly attributed to 1) the availability of large amounts of in-domain development data from a large number of labeled speakers, 2) speaker representations (aka embeddings) extracted using extended and more complex end-to-end neural network frameworks, and 3) effective use of the provided large development set.

In 2019, the U.S. National Institute of Standards and Technology (NIST) conducted the most recent in an ongoing series of speaker recognition evaluations (SRE). There were two components to SRE19: 1) a leaderboard style Challenge using unexposed conversational telephone speech (CTS) data from the Call My Net 2 (CMN2) corpus, and 2) an Audio-Visual (AV) evaluation using video material extracted from the unexposed portions of the Video Annotation for Speech Technologies (VAST) corpus. This paper presents an overview of the Audio-Visual SRE19 activity including the task, the performance metric, data, and the evaluation protocol, results and system performance analyses. The Audio-Visual SRE19 was organized in a similar manner to the audio from video (AfV) track in SRE18, except it offered only the open training condition. In addition, instead of extracting and releasing only the AfV data, unexposed multimedia data from the VAST corpus was used to support the Audio-Visual SRE19. It featured two core evaluation tracks, namely audio only and audio-visual, as well as an optional visual only track. A total of 26 organizations (forming 14 teams) from academia and industry participated in the Audio-Visual SRE19 and submitted 102 valid system outputs. Evaluation results indicate: 1) notable performance improvements for the audio only speaker recognition task on the challenging amateur online video domain due to the use of more complex neural network architectures (e.g., ResNet) along with soft margin losses, 2) state-of-the-art speaker and face recognition technologies provide comparable person recognition performance on the amateur online video domain, and 3) audio-visual fusion results in remarkable performance gains (greater than 85% relative) over the audio only or visual only systems.

The Director's Office at MIT Lincoln Laboratory (MIT LL) requested a comprehensive study on artificial intelligence (AI) focusing on present applications and future science and technology (S&T) opportunities in the Cyber Security and Information Sciences Division (Division 5). This report elaborates on the main results from the study. Since the AI field is evolving so rapidly, the study scope was to look at the recent past and ongoing developments to lead to a set of findings and recommendations. It was important to begin with a short AI history and a lay-of-the-land on representative developments across the Department of Defense (DoD), intelligence communities (IC), and Homeland Security. These areas are addressed in more detail within the report. A main deliverable from the study was to formulate an end-to-end AI canonical architecture that was suitable for a range of applications. The AI canonical architecture, formulated in the study, serves as the guiding framework for all the sections in this report. Even though the study primarily focused on cyber security and information sciences, the enabling technologies are broadly applicable to many other areas. Therefore, we dedicate a full section on enabling technologies in Section 3. The discussion on enabling technologies helps the reader clarify the distinction among AI, machine learning algorithms, and specific techniques to make an end-to-end AI system viable. In order to understand what is the lay-of-the-land in AI, study participants performed a fairly wide reach within MIT LL and external to the Laboratory (government, commercial companies, defense industrial base, peers, academia, and AI centers). In addition to the study participants (shown in the next section under acknowledgements), we also assembled an internal review team (IRT). The IRT was extremely helpful in providing feedback and in helping with the formulation of the study briefings, as we transitioned from datagathering mode to the study synthesis. The format followed throughout the study was to highlight relevant content that substantiates the study findings, and identify a set of recommendations. An important finding is the significant AI investment by the so-called "big 6" commercial companies. These major commercial companies are Google, Amazon, Facebook, Microsoft, Apple, and IBM. They dominate in the AI ecosystem research and development (R&D) investments within the U.S. According to a recent McKinsey Global Institute report, cumulative R&D investment in AI amounts to about $30 billion per year. This amount is substantially higher than the R&D investment within the DoD, IC, and Homeland Security. Therefore, the DoD will need to be very strategic about investing where needed, while at the same time leveraging the technologies already developed and available from a wide range of commercial applications. As we will discuss in Section 1 as part of the AI history, MIT LL has been instrumental in developing advanced AI capabilities. For example, MIT LL has a long history in the development of human language technologies (HLT) by successfully applying machine learning algorithms to difficult problems in speech recognition, machine translation, and speech understanding. Section 4 elaborates on prior applications of these technologies, as well as newer applications in the context of multi-modalities (e.g., speech, text, images, and video). An end-to-end AI system is very well suited to enhancing the capabilities of human language analysis. Section 5 discusses AI's nascent role in cyber security. There have been cases where AI has already provided important benefits. However, much more research is needed in both the application of AI to cyber security and the associated vulnerability to the so-called adversarial AI. Adversarial AI is an area very critical to the DoD, IC, and Homeland Security, where malicious adversaries can disrupt AI systems and make them untrusted in operational environments. This report concludes with specific recommendations by formulating the way forward for Division 5 and a discussion of S&T challenges and opportunities. The S&T challenges and opportunities are centered on the key elements of the AI canonical architecture to strengthen the AI capabilities across the DoD, IC, and Homeland Security in support of national security.

Recent work has shown large performance gains using denoising DNNs for speech processing tasks under challenging acoustic conditions. However, training these DNNs requires large amounts of parallel multichannel speech data which can be impractical or expensive to collect. The effective use of synthetic parallel data as an alternative has been demonstrated for several speech technologies including automatic speech recognition and speaker recognition (SR). This paper demonstrates that denoising DNNs trained with real Mixer 2 multichannel data perform only slightly better than DNNs trained with synthetic multichannel data for microphone SR on Mixer 6. Large reductions in pooled error rates of 50% EER and 30% min DCF are achieved using DNNs trained on real Mixer 2 data. Nearly the same performance gains are achieved using synthetic data generated with a limited number of room impulse responses (RIRs) and noise sources derived from Mixer 2. Using RIRs from three publicly available sources used in the Kaldi ASpIRE recipe yields somewhat lower pooled gains of 34% EER and 25% min DCF. These results confirm the effective use of synthetic parallel data for DNN channel compensation even when the RIRs used for synthesizing the data are not particularly well matched to the task.

Over several decades, speaker recognition performance has steadily improved for applications using telephone speech. A big part of this improvement has been the availability of large quantities of speaker-labeled data from telephone recordings. For new data applications, such as audio from room microphones, we would like to effectively use existing telephone data to build systems with high accuracy while maintaining good performance on existing telephone tasks. In this paper we compare and combine approaches to compensate models parameters and features for this purpose. For model adaptation we explore MAP adaptation of hyper-parameters and for feature compensation we examine the use of denoising DNNs. On a multi-room, multi-microphone speaker recognition experiment we show a reduction of 61% in EER with a combination of these approaches while slightly improving performance on telephone data.

In this paper we describe the most recent MIT Lincoln Laboratory language recognition system developed for the NIST 2015 Language Recognition Evaluation (LRE). The submission features a fusion of five core classifiers, with most systems developed in the context of an i-vector framework. The 2015 evaluation presented new paradigms. First, the evaluation included fixed training and open training tracks for the first time; second, language classification performance was measured across 6 language clusters using 20 language classes instead of an N-way language task; and third, performance was measured across a nominal 3-30 second range. Results are presented for the overall performance across the six language clusters for both the fixed and open training tasks. On the 6-cluster metric the Lincoln system achieved overall costs of 0.173 and 0.168 for the fixed and open tasks respectively.

The development of the i-vector framework for generating low dimensional representations of speech utterances has led to considerable improvements in speaker recognition performance. Although these gains have been achieved in periodic National Institute of Standards and Technology (NIST) evaluations, the problem of domain mismatch, where the system development data and the application data are collected from different sources, remains a challenging one. The impact of domain mismatch was a focus of the Johns Hopkins University (JHU) 2013 speaker recognition workshop, where a domain adaptation challenge (DAC13) corpus was created to address this problem. This paper proposes an approach to domain mismatch compensation for applications where in-domain development data is assumed to be unavailable. The method is based on a generalization of data whitening used in association with i-vector length normalization and utilizes a library of whitening transforms trained at system development time using strictly out-of-domain data. The approach is evaluated on the 2013 domain adaptation challenge task and is shown to compare favorably to in-domain conventional whitening and to nuisance attribute projection (NAP) inter-dataset variability compensation.

Significant performance gains have been reported separately for speaker recognition (SR) and language recognition (LR) tasks using either DNN posteriors of sub-phonetic units or DNN feature representations, but the two techniques have not been compared on the same SR or LR task or across SR and LR tasks using the same DNN. In this work we present the application of a single DNN for both tasks using the 2013 Domain Adaptation Challenge speaker recognition (DAC13) and the NIST 2011 language recognition evaluation (LRE11) benchmarks. Using a single DNN trained on Switchboard data we demonstrate large gains in performance on both benchmarks: a 55% reduction in EER for the DAC13 out-of-domain condition and a 48% reduction in Cavg on the LRE11 30s test condition. Score fusion and feature fusion are also investigated as is the performance of the DNN technologies at short durations for SR.

The impressive gains in performance obtained using deep neural networks (DNNs) for automatic speech recognition (ASR) have motivated the application of DNNs to other speech technologies such as speaker recognition (SR) and language recognition (LR). Prior work has shown performance gains for separate SR and LR tasks using DNNs for direct classification or for feature extraction. In this work we present the application for single DNN for both SR and LR using the 2013 Domain Adaptation Challenge speaker recognition (DAC13) and the NIST 2011 language recognition evaluation (LRE11) benchmarks. Using a single DNN trained for ASR on Switchboard data we demonstrate large gains on performance in both benchmarks: a 55% reduction in EER for the DAC13 out-of-domain condition and a 48% reduction in Cavg on the LRE11 30 s test condition. It is also shown that further gains are possible using score or feature fusion leading to the possibility of a single i-vector extractor producing state-of-the-art SR and LR performance.

We consider the use of community detection algorithms to perform speaker clustering on content graphs built from large audio corpora. We survey the application of agglomerative hierarchical clustering, modularity optimization methods, and spectral clustering as well as two random walk algorithms: Markov clustering and Infomap. Our results on graphs built from the NIST 2005+2006 and 2008+2010 Speaker Recognition Evaluations (SREs) provide insight into both the structure of the speakers present in the data and the intricacies of the clustering methods. In particular, we introduce an additional parameter to Infomap that improves its clustering performance on all graphs. Lastly, we also develop an automatic technique to purify the neighbors of each node by pruning away unnecessary edges.