ONLINE SPEAKER DIARIZATION USING ADAPTED I-VECTOR TRANSFORMS Weizhong Zhu and Jason Pelecanos IBM Research, Yorktown Heights, NY 1598, USA {zhuwe,jwpeleca}@us.ibm.com ABSTRACT Many speaker diarization systems operate in an off-line mode. Such systems typically find homogeneous segments and then cluster these segments according to speaker. Such algorithms, like bottom-up clustering, k-means or spectral clustering, generally require the registration of all segments before clustering can begin. However, for real-time applications such as with multi-person voice interactive systems, there is a need to perform online speaker assignment in a strict left-toright fashion. In this paper we propose a novel Maximum a Posteriori (MAP) adapted transform within an i-vector speaker diarization framework, that operates in a strict left-to-right fashion. Previous work by the community has shown that the principal components of variation of fixed dimensional i-vectors learned across segments tend to indicate a strong basis by which to separate speakers. However, determining this basis can be problematic when there are few segments or when operating in an online manner. The proposed method blends the prior with the estimated subspace as more i-vectors are observed. Given oracle SAD segments, with adaptation we achieve 3.2% speaker diarization error for a strict left-to-right constraint on the LDC Callhome English Corpus compared to 4.8% without adaptation. Index Terms i-vector, speaker recognition, speaker diarization, left-to-right clustering 1. INTRODUCTION Many publications on speaker diarization have considered the problem from an off-line mode perspective, where the entire recording may be processed before the labeling decisions are made [1, 2]. In this paper, we consider the challenge of performing speaker diarization where the system is required to make a concrete labeling decision at each point in time once a change point is detected. This requirement is important in many streaming based applications. For conversational speech this can be useful for improving online speech recognition system performance by tracking speaker specific parameters over time. It is also useful for providing speaker recognition capabilities to support multi-person human-computer voice interactive systems. Past work for on-line speaker diarization systems has involved the use of Gaussian Mixture Models [3, 4, 5] and hybrid online-offline mechanisms using MAP adapted GMMs [6]. In this work, instead of directly scoring segments represented by GMMs, we perform diarization based on the i-vector representation of acoustic segments. The i-vector approach has become a mainstream front-end in state-of-the-art speaker [7] and language recognition systems and more recently in offline diarization systems [8, 9]. For a typical speaker diarization system, a speech activity detector and change point detector partition the recording into short segments. Such a system could partition segments based on a homogeneity criterion such as the Bayesian Information Criterion (BIC) [1] or based on segments derived from speech detection systems [11, 12]. Once these segments are determined, a fixed dimension i-vector representation is then calculated for each segment in the recording. The i-vectors are then compared (using PLDA [13] or a cosine distance metric) and merged according to an agglomerative clustering schedule for offline systems or by incremental comparison for online systems. In the work of Shum [8] it was observed that the i-vectors across segments of a recording varied in a manner such that the first few principal components of variation contained significant information pertaining to the different speakers in the recording. Shum and [9] have shown that significant performance improvement can be achieved in offline systems by clustering using these first few dimensions. In contrast to previous work, we modify this approach to target online analysis and we accordingly incorporate a MAP based formulation to adjust the speaker space as more segments are observed in the recording. The remainder of the paper is organized as follows: Section 2 gives some insight into the nature of i-vectors in the experimental speaker diarization task. Section 3 describes the proposed MAP adapted subspace estimation algorithm. Section 4 describes the experimental setup and results and Section 5 wraps up with the conclusions. 2. I-VECTOR PRINCIPAL COMPONENTS: A VISUALIZATION We further explore the work by Shum [8] that suggests using the principal components describing the variation of i-vectors 978-1-4799-9988-/16/$31. 216 IEEE 545 ICASSP 216
the female-female conversation tends to have shorter more rapid-fire turns while the male-female conversation tends to have longer turns. The plots also show that the segments that are more difficult to differentiate are shorter in duration while the i-vectors based on longer segments are better separated. As past experimental research has demonstrated [8], and as these two-speaker telephone conversation examples support, the principal components of i-vector variation play an important role in differentiating between speakers. Given that this is a successful approach for offline analysis, the challenge now is to modify the i-vector principal components approach for online diarization. 1.8 Fig. 1. Histogram of the segment durations based on the transcriptions of the recordings from the LDC CALLHOME English corpus. PCA2.6.4.2 of segments to separate speakers. In this section, we look into some recordings and their corresponding distributions of i- vectors, so we can gain further insight into the principal component method. To better understand the nature of the diarization i- vectors, we explore two-speaker conversations from the LDC CALLHOME English corpus [14]. For each recording, an i-vector is calculated for each oracle SAD segment within the recording. For this analysis and in the experiments, the oracle segments are based on the reference human transcriptions. Each conversation is about 8-1 minutes long. Figure 1 shows the histogram of segment durations from all conversations. There is a peak near the.4 second bin and, after the peak, the counts decay rapidly as duration increases. Although some segments can be more than 1 seconds in duration, most segments are relatively short. The average segment duration is 2.1 seconds. To demonstrate the separation provided by the principal components of i-vectors, in Figure 2 we present a visualization of segments for two principal components. Each dot or filled circle in the figure represents an i-vector based segment with the radius of the dot representing the duration of the segment. The two different colors represent two different speakers. Figure 2(a) shows the result for a female-female conversation while Figure 2(b) presents a male-female conversation. As can be observed from the figures, i-vectors from two different speakers tend to form two major clouds; one to the left and the other to the right side of the x-axis. The x-axis direction, which relates to the first principal component, is indicative of a strong direction of discrimination between the two speakers. We also observe that the size of the circles indicates that PCA2.2.4.6.8 1 1.8.6.4.2.2.4.6.8 1 PCA1 1.8.6.4.2.2.4.6.8 (a) Female-Female Conversation 1 1.8.6.4.2.2.4.6.8 1 PCA1 (b) Male-Female Conversation Fig. 2. Plots of the first two principal components of the i- vectors estimated from the segments of each two-speaker telephone conversation (a) and (b). 546
3. INCORPORATING THE SUBSPACE APPROACH FOR ONLINE ANALYSIS As identified in the previous section, the principal component subspace used to separate speakers is formed from the entire set of i-vectors from the recording. If we assert that the diarization process must be operated in strict left-to-right fashion, we need a modified approach. Here, we propose the following algorithm to estimate an i-vector transform (rather than a subspace) that is adapted as more i-vectors are registered. T n = α n V n V n + (1 α n )I (1) α n = n n + R The variable V n, represents the first principal component (eigenvector) estimated from all segments currently accumulated up until segment number n. (We note that additional principal components can be included for discriminating between more than two speakers.) The identity matrix is indicated as I and ( ) is the transpose. The matrix, T n, is the i-vector transformation matrix where an adaptation factor α smoothes the estimated speaker subspace with the identity matrix. The rate at which the system adapts is controlled by the relevance factor, R. The formulation for T n is based on a MAP adaptation schedule [15, 16, 17]. In addition to the assumptions made in these publications, this formulation results when the mean of the prior distribution of the i-vectors and the adapted mean are set to. We note that other formulations of this adaptation approach are possible but are not explored here. There are particular cases that may be represented by this formulation. When the relevance factor R is set to infinity, the transformation becomes the identity matrix and the i-vectors are not transformed. When the relevance factor is zero, only the subspace calculated up to the current point in time is used. i.e. no prior information is used. If a more reasonable relevance factor is specified (for example, R = 16), before the first segment is acquired, the transformation begins as the identity matrix. As more segments are acquired, more emphasis is placed on the subspace specified by the segment based i-vectors. In the limit, as the number of segments tends to infinity, the system no longer relies on the prior information to determine the transformation. Given the formulation for the transformation T n, we now summarize the procedure for a strict left-to-right diarization system. The first step is to identify a new segment 1. Once the i-vector for this segment is calculated, the task is then to assign the segment to a new cluster or to one of the existing clusters. For the nth new segment, we first use equation 1 1 There are many ways to achieve this including BIC analysis [1] and SAD segmentation [11]. In our work we use to oracle SAD segments. to estimate the transformation T n and then we apply this estimated transform to all observed i-vectors. The cosine distances are calculated between all i-vectors from the previous segments and the nth i-vector to give n 1 scores. For each cluster, the scores of the segments associated with each cluster are averaged. If the best scoring cluster has a score larger than a threshold then the nth i-vector is assigned to the cluster, otherwise a new cluster is allocated. This process continues for each new segment. 4.1. Data and System 4. EXPERIMENTS In this section we describe the experimental setup used in our evaluations. We begin by describing the training material and then provide a brief overview of the i-vector extraction process. For training the system, speech material is drawn from the NIST 5, 6, 8, 1 speaker recognition evaluations and the Switchboard English telephone conversation corpus [18, 19]. The training set consists of 3k recordings and has a total of 419 speakers. We extracted a total of 48k segments from these recordings. All extracted segments are limited to greater than 3 seconds based on a speech activity detection (SAD) component that uses energy, voicing, and spectral divergence parameters [2]. The test data is comprised of 113 two-speaker conversations from the LDC CALLHOME English corpus [14]. We use oracle SAD segments from the human transcriptions. Each conversation is about 1 minutes long. The system configuration is briefly described. For speech parameterization, 19-dimensional MFCCs (32ms frames every 1ms using a 24-channel Mel-filterbank) span the frequency range of 125-37Hz. The MFCC features do not have channel compensation applied. To learn the i-vector extractor, a 124-component gender-independent diagonal GMM is trained. The dimensionality of the total variability subspace is set to 64. The Within Class Covariance Normalization (WCCN) [21] matrix is estimated by using the aforementioned training material and is applied to all i-vectors prior to applying the adaptive transformation, T n. In contrast to the typical WCCN formulation for speaker recognition, the speaker specific statistics are calculated at the short segment level rather than the session level. 4.2. Results and Discussion Before discussing the results, we outline the error metric used. A common metric, called the Speaker Diarization Error Rate (DER) is the sum of 3 types of errors: 1. Speaker Miss Error: percentage of time that the speaker labels are not assigned to speech; 2. Speaker False Alarm Error: percentage of time that the speaker labels are assigned falsely to non-speech; 3. Speaker Error: percentage of time that the speaker labels are 547
Speaker Error (%) Table 1. The speaker diarization performance with oracle SADs by using i-vectors only, i-vectors with the WCCN transform and with two different clustering methods. (The subspace is estimated using all segments from the recording.) DER (%) DER (%) Method of Clustering i-vectors i-vectors+wccn Bottom-up (Offline) 2.4 2.1 Left-to-right (Online) 3.7 3.2 assigned to the wrong speaker. We used oracle SADs for the following experiments, so the error resolves to being simply speaker labeling error. (Following the conventions for evaluating diarization performance [22], the evaluation code ignores intervals containing overlapped speech and tolerates errors with 25ms segment boundaries.) Our experiments begin by presenting the results where the subspace is estimated using all i-vectors from a recording and regular bottom-up clustering is performed (or effectively the offline mode). As a comparison, we keep the estimated subspace and then perform left-to-right clustering based only on this subspace. The promising left-to-right results demonstrate the importance of finding a reliable and speaker discriminative subspace for online operation. Table 1 also shows results for these two scenarios with and without WCCN i- vector compensation applied. WCCN provides about 1% relative performance improvement for both bottom-up and left-to-right clustering. Now we present results for the same test set labeled in a strict left-to-right way without knowing the subspace ahead of time. We use equation (1) to estimate the i-vector transform, apply the estimated transform to the currently observed i-vectors and perform classification sequentially on each segment from left-to-right. Table 2 shows the speaker error rate as we vary the relevance factor R. It is apparent that the algorithm can be operated robustly across a relatively broad range. The results show that the performance of the proposed technique is relatively good for long recordings. In order to understand how the different durations of the recording affect performance, we examine the speaker error rate as we limit the length of the recording to evaluate. The results are in Figure 3. The x-axis represents the duration of the recording to evaluate, and the y-axis is the speaker error. The solid line represents the proposed algorithm with R = 16, while the dashed line represents no adaptation, where R =. As expected, for short recordings the impact of the transform is small. However, for longer durations the benefit is more pronounced; up to 37% relative error reduction is achieved. Table 2. Online speaker diarization performance using a MAP adapted transform with different relevance factors R. 12 1 8 6 4 2 Relevance Factor Speaker Error Rate (%) 1. 1 5. 2 4.6 4 3.9 8 3.6 16 3.2 32 3.3 64 3.4 128 3.8 256 4.3 512 4.5 124 4.8 4.8 Comparison of Diarization Performance vs Recording Duration 1 2 3 4 5 6 7 8 9 1 Duration of Recording (Minutes) Proposed Method (R=16) No Adaptation (R= ) Fig. 3. A plot of performance vs. recording duration for leftto-right diarization. 5. CONCLUSIONS We proposed a novel MAP based i-vector transformation method and applied it to the problem of online speaker diarization. We have shown that, for two speaker conversation data from the LDC English CALLHOME corpus, the proposed technique can achieve 3.2% speaker error. As compared with the standard method with no adaptation, we observe an error reduction of about one third when the duration of the recording is longer than 2 minutes. We also show that performance can be improved by applying a specialized configuration of WCCN to the i-vectors. 548
6. REFERENCES [1] S. E. Tranter and D. A. Reynolds, An overview of automatic speaker diarization system, in IEEE Transactions on Audio, Speech, and Language Processing, 26. [2] X. A. Miro et al., Speaker diarization: A review of recent research, in IEEE Transactions on Audio, Speech, and Language Processing, 212. [3] K. Markov and S. Nakamura, Never-ending learning system for on-line speaker diarization, in IEEE Automatic Speeach Recognition and Understanding Workshop, 27. [4] J. Geiger, F. Wallhoff, and G. Rigoll, GMM-UBM based open-set online speaker diarization, in Inter- Speech, 21. [5] G. Soldi, C. Beaugeant, and N. Evans, Adaptive and online speaker diarization for meeting data, in EU- SIPCO, 215. [6] C. Vaquero, O. Vinyals, and G. Friedland, A hybrid approach to online speaker diarization, in InterSpeech, 21. [7] N. Dehak et al., Front-end factor analysis for speaker verification, IEEE Transactions on Audio, Speech, and Language Processing, 211. [8] S. H. Shum et al., Unsupervised methods for speaker diarization: An integrated and iterative approach, in IEEE Transactions on Audio, Speech, and Language Processing, 213. [14] A. Canavan et al., CALLHOME American English Speech, 1997. [15] D. Reynolds, T. Quatieri, and R. Dunn, Speaker verification using adapted Gaussian mixture models, Digital Signal Processing, 2. [16] J. L. Gauvain and C-H. Lee, Maximum a posteriori estimation for multivariate Gaussian mixture observations of Markov chains, in IEEE Transactions on Speech and Audio Processing, 1994. [17] J. Pelecanos, Robust automatic speaker recognition, Ph.D. Thesis, Queensland University of Technology, 23. [18] D. Graff et al., Switchboard-2 Phase II, Linguistic Data Consortium, 1999. [19] J. Godfrey et al., Switchboard-2 Phase III audio, Linguistic Data Consortium, 22. [2] W. Zhu et al., Nearest neighbor based i-vector normalization for robust speaker recognition under unseen channel conditions, in International Conference on Acoustics, Speech, and Signal Processing, 215. [21] A. Hatch, S. Kajarekar, and A. Stolcke, Withinclass covariance normalization for SVM-based speaker recognition, in International Conference on Spoken Language Processing, 26. [22] NIST, Rich transcription evaluation project, http://www.itl.nist.gov/iad/mig/tests/rt/, 215. [9] D. Garcia-Romero and A. McCree, Speaker diarization with PLDA i-vector scoring and unsupervised calibration, in IEEE Spoken Language Technology Workshop, 215. [1] S. S. Chen and P. S. Gopalakrishnan, Speaker, environment and channel change detection and clustering via the Bayesian information criterion, in DARPA Broadcast News Transcription and Understanding Workshop, 1998. [11] M. A. Siegler et al., Automatic segmentation, classification and clustering of broadcast news audio, in DARPA Speech Recognition Workshop, 1997. [12] D. Haws et al., On the importance of event detection for ASR, in International Conference on Acoustics, Speech, and Signal Processing, 216. [13] S. J. D. Prince and J. H. Elder, Probabilistic linear discriminant analysis for inferences about identity, in International Conference on Computer Vision, 27. 549