International audience

ASVspoof challenges are designed to advance the understanding of spoofing speech attacks and encourage the development of robust countermeasure systems. These challenges provide a standardized database for assessing and comparing spoofing-robust automatic speaker verification solutions. The ASVspoof5 challenge introduces a shift in database conditions compared to ASVspoof2019. While ASVspoof2019 has mismatched conditions only in spoofing attacks in the evaluation set, ASVspoof5 incorporates mismatches in both bona fide and spoofed speech statistics. This paper examines the impact of these mismatches, presenting qualitative and quantitative comparisons within and between the two databases. We show the increased difficulty for genuine and spoofed speech and demonstrate that in ASVspoof5, not only are the attacks more challenging, but the genuine speech also shifts toward spoofed speech compared to ASVspoof2019.

Anti-spoofing is the task of speech authentication. That is, identifying genuine human speech compared to spoofed speech. The main focus of this paper is to suggest new representations for genuine and spoofed speech, based on the probability mass function (PMF) estimation of the audio waveforms' amplitude. We introduce a new feature extraction method for speech audio signals: unlike traditional methods, our method is based on direct processing of time-domain audio samples. The PMF is utilized by designing a feature extractor based on different PMF distances and similarity measures. As an additional step, we used filterbank preprocessing, which significantly affects the discriminative characteristics of the features and facilitates convenient visualization of possible clustering of spoofing attacks. Furthermore, we use diffusion maps to reveal the underlying manifold on which the data lies. The suggested embeddings allow the use of simple linear separators to achieve 12.99% Equal Error Rate (EER) on ASVspoof2019 logical Access (LA) test set for female samples, and 12.09% for male samples. In addition, we present a convenient way to visualize the data, which helps to assess the efficiency of different spoofing techniques. Furthermore, we present reduced complexity embedding method by using compander quantization, which in some cases even improves the EER on the test set up to 3.00%. The experimental results show the potential of using multichannel PMF-based features for the anti-spoofing task, in addition to the benefits of using diffusion maps both as an analysis tool and as an embedding tool.

Speaker Change Detection (SCD) is the task of segmenting an input audio-recording according to speaker interchanges. Nowadays, many applications, such as Speaker Diarization (SD) or automatic vocal transcription, depend on this segmentation task. In this paper, we focus on the essential task of the SD problem, the audio segmenting process, and suggest a solution for the SCD problem, as well as the assignment of clustered speaker labels for the extracted segments, and applying the solution over two datasets: a commercial dataset in Hebrew and the ICSI Meeting Corpus. As such, we propose a hybrid framework for the SCD problem that is learned by textual information and speech signals and the meta-data features that can be extracted from them. Moreover, we demonstrate the negative correlation between an increase in the number of speakers in the training dataset and the influence on the overall diarization system's performance, which is improved using our efficient SCD component. Finally, we show how our proposed hybrid framework remains robust compared to the ICSI Meeting Corpus, as the experimental evaluation's training and testing is based on two languages.

International audience

Speaker diarization systems attempt to assign temporal segments from a conversation betweenRspeakers to an appropriate speakerr. This task is generally performed when no prior information is given regarding the speakers. The number of speakers is usually unknown and needs to be estimated. However, there are applications where the number of speakers is known in advance. The diarization process generally consists of change detection, clustering and labeling of a given audio stream. Speaker diarization can be performed using an iterative approach that is optimized by the selection of appropriate initial conditions. This study examines the influence of several common initialization algorithms including two variants of a recently proposed, K-means based initialization algorithm over the performance of an iterative-based speaker diarization system applied to two speaker telephone conversations. The suggested speaker diarization system employs either self organizing maps or Gaussian mixture models in order to model the speakers and non-speech in the conversation. The diarization system and initialization algorithms are tuned using 108 telephone conversations taken from LDC CallHome corpus, this is the development set. The evaluation subset is composed of 2048 telephone conversations extracted from the NIST 2005 Rich Transcription corpus. The results obtained show that by initializing the speaker diarization system using the K-means based algorithms provide a relative improvement of 10.4% for the LDC development set and 12.2% for the NIST evaluation subset when compared to random initialization after 12 iterations which are required for the convergence of the diarization process using random initialization. However, when using the K-means based initialization approach, only five iterations are required for the system to converge. Thus, using the new initialization allows us to improve the performances both in terms of diarization error rate and speed of convergence.

Speaker diarization systems attempt segmentation and labeling of a conversation between R speakers, while no prior information is given regarding the conversation. Most state of the art diarization systems require the full body of the conversation data prior to the application of some diarization approach. However, for some applications such as forensics, which handles vast amount of data, an on-line or incremental diarization is of high importance. For that purpose, a two-stage incremental diarization of telephone conversations algorithm is suggested. On the first stage, a fully unsupervised diarization algorithm is applied over an initial training segment from the conversation. The secondstage is composed of time-series clustering of increments of the conversation. Applying incremental diarization over 1802 telephone conversations from NIST 2005 SER generated an increase in diarization error of approximately 2% compared to the diarization error of an off-line diarization system.

One inherent deficiency of most diarization systems is their inability to handle co-channel or overlapped speech. Most of the suggested algorithms perform under singular conditions, require high computational complexity in both time and frequency domains. In this study, frame based entropy analysis of the audio data in the time domain serves as a single feature for an overlapped speech detection algorithm. Identification of overlapped speech segments is performed using Gaussian Mixture Modeling (GMM) along with well known classification algorithms applied on two speaker conversations. By employing this methodology, the proposed method eliminates the need for setting a hard threshold for each conversation or database. LDC CALLHOME American English corpus is used for evaluation of the suggested algorithm. The proposed method successfully detects 63.2% of the frames labeled as overlapped speech by the manual segmentation, while keeping a 5.4% false-alarm rate.

A new approach for initial assignment of data in a speaker clustering application is presented. This approach employs Weighted Segmental K-Means clustering algorithm prior to competitive based learning. The clustering system relies on Self-Organizing Maps (SOM) for speaker modeling and likelihood estimation. Performance is evaluated on 108 two speaker conversations taken from LDC CALLHOME American English Speech corpus using NIST criterion and shows an improvement of approximately 48% in Cluster Error Rate (CER) relative to the randomly initialized clustering system. The number of iterations was reduced significantly, which contributes to both speed and efficiency of the clustering system.

A new approach is presented for clustering the speakers from unlabeled and unsegmented conversation, when the number of speakers is unknown. In this approach, Self-Organizing-Map (SOM) is used as likelihood estimators for speaker model. For estimation of the number of clusters the Bayesian Information Criterion (BIC) is applied. This approach was tested on the NIST 1996 HUB-4 evaluation test in terms of speaker and cluster purities. Results indicate that the combined SOM-BIC approach can lead to better clustering results than the baseline system. 1.

In many time-series such as speech, biosignals, protein chains, etc. there is a dependency between consecutive vectors. As the dependency is limited in duration, such data can be referred to as piecewise-dependent data (PDD). In clustering, it is frequently needed to minimize a given distance function. In this letter, we will show that in PDD clustering there is a contradiction between the desire for high resolution (short segments and low distance) and high accuracy (long segments and high distance), i.e., meaningful clustering.

We present a method for clustering the speakers from unlabeled and unsegmented conversation (with known number of speakers), when no a priori knowledge about the identity of the participants is given. Each speaker was modeled by a self-organizing map (SOM). The SOMs were randomly initiated. An iterative algorithm allows the data move from one model to another and adjust the SOMs. The restriction that the data can move only in small groups but not by moving each and every feature vector separately force the SOMs to adjust to speakers (instead of phonemes or other vocal events). This method was applied to high-quality conversations with two to five participants and to two-speaker telephone-quality conversations. The results for two (both high- and telephone-quality) and three speakers were over 80% correct segmentation. The problem becomes even harder when the number of participants is also unknown. Based on the iterative clustering algorithm a validity criterion was also developed to estimate the number of speakers. In 16 out of 17 conversations of high-quality conversations between two and three participants, the estimation of the number of the participants was correct. In telephone-quality the results were poorer.