Overlapped speech detection for improved speaker diarization in multiparty meetings

Single ion channel signal of cell membrane is a stochastic ionic currents in the order of picoampere (pA). Because of the weakness of the signal, the background noise always dominates in the patch-clamp recordings. The threshold detector is traditionally used to eliminate noise and restore the single channel signal. However, this method cannot work satisfactorily when signal-to-noise ratio is lower. An approach based on hidden Markov model (HMM) is used to reconstruct ionic single-channel currents and estimate model parameters under white background noise. Firstly, ionic single-channel currents were depicted and analyzed by HMM. Then, an iterative algorithm of Baum-Welch was introduced to train HMM and estimate the model parameters. Finally, the ideal channel currents were reconstructed by Viterbi algorithm. Compared HMM with the threshold detection by computer simulating under different transition probabilities and signal-to-noise ratios, and the results have shown that the method performs effectively under the low signal-to-noise ratio (SNR<;5.0) and has fast model parameter convergence, high restoration precision and strong noise robustness.

In this work, we perform a first approach to emotion recognition from EEG single channel signals extracted in four (4) mother-child dyads experiment in developmental psychology. Single channel EEG signals are analyzed and processed using several window sizes by performing a statistical analysis over features in the time and frequency domains. Finally, a neural network obtained an average accuracy rate of 99% of classification in two emotional states such as happiness and sadness.

The electroencephalogram (EEG) signals often contaminated by muscle or electromyogram (EMG) artifacts. The presence of these artifacts obscure the desired information in the EEG signal. In this paper, we proposed an efficient subspace based technique named singular spectrum analysis (SSA), to remove the EMG artifacts from the single channel EEG signals. For this we proposed a new grouping technique to extract efficiently the desired component from the contaminated EEG signal by setting a threshold. Firstly, single channel signal is mapped into multichannel signal or data, called embedding. Next, the orthogonal eigenvectors are estimated from the covariance matrix of the multichannel data by singular value decomposition (SVD). Since the local variations of eigenvectors corresponding to the EEG signals are low compare with the EMG signals, we set an arbitrary threshold (0.275) to find these eigenvectors, which are used to create the subspace corresponding to the EEG signals. After identifying the subspace, the EEG signals are extracted by simply projecting the multichannel data onto this subspace followed by reverse process of embedding step. Finally, the proposed method is applied on synthetic noisy sinusoidal signals and EEG signals contaminated by the EMG artifacts. The results shows that the proposed method can efficiently removes the EMG artifacts without altering the desired components.

The performance of most speaker diarization systems with x-vector embeddings is both vulnerable to noisy environments and lacks domain robustness. Earlier work on speaker diarization using generative adversarial network (GAN) with an encoder network (ClusterGAN) to project input x-vectors into a latent space has shown promising performance on meeting data. In this paper, we extend the ClusterGAN network to improve diarization robustness and enable rapid generalization across various challenging domains. To this end, we fetch the pre-trained encoder from the ClusterGAN and fine tune it by using prototypical loss (meta-ClusterGAN or MCGAN) under the meta-learning paradigm. Experiments are conducted on CALLHOME telephonic conversations, AMI meeting data, DIHARD-II (dev set) which includes challenging multi-domain corpus, and two child-clinician interaction corpora (ADOS, BOSCC) related to the autism spectrum disorder domain. Extensive analyses of the experimental data are done to investigate the effectiveness of the proposed ClusterGAN and MCGAN embeddings over x-vectors. The results show that the proposed embeddings with normalized maximum eigengap spectral clustering (NME-SC) back-end consistently outperform the Kaldi state-of-the-art x-vector diarization system. Finally, we employ embedding fusion with x-vectors to provide further improvement in diarization performance. We achieve a relative diarization error rate (DER) improvement of 6.67% to 53.93% on the aforementioned datasets using the proposed fused embeddings over x-vectors. Besides, the MCGAN embeddings provide better performance in the number of speakers estimation and short speech segment diarization compared to x-vectors and ClusterGAN on telephonic conversations.

As one kind of current main power sources, internal combustion engines require high-reliability to ensure mechanical systems working well in normal operation. This paper studies vibration signals in order to detect multiple types of faults by a single channel signal. First, for the decomposition level of variational mode decomposition (VMD) which needs to be chosen non-automatically, this paper analyzes the features of various faults and optimizes the iteration initial values of center frequency so as to reduce the adverse effect of inappropriate decomposition level. Moreover, considering that VMD cannot decompose different signal sources in the same frequency, robust independent component analysis (ICA) is an excellent method to overcome this challenge. Then, the fourth-order cumulant of restructured signals from VMD and robust ICA is taken as fault indexes. Finally, because of the high error rate of original fuzzy C-Mean clustering, Euclidean distance between test points and cluster center of fuel injection failure is taken as a fault detection threshold in order to achieve high-recognition rate. In conclusion, through optimizing several algorithms, the purpose of detecting multiple types of faults by single channel signal is achieved in the current research.

Feature mode decomposition (FMD) has advantages over the other newer time-frequency methods such as ensemble empirical mode decomposition (EEMD) and variational mode decomposition (VMD) in single channel blind source separation (SCBSS). However, FMD has the defect of needing to determine the precise number of fault sources manually. To solve the above defect of FMD, an adaptive method for determining the number of fault sources based on the shift invariant sparse code (SISC) is proposed. SISC was used to train a set of basis functions from the single channel signal, and the corresponding potential components were reconstructed firstly. Subsequently, the structural similarity of these potential components was used for clustering, and each of the obtained clustering signals represented one kind of fault. Then the number of clustering was determined by minimizing the structural correlation among the clustering signals. It was considered that the source separation had achieved the best effect when the structural difference among the clusters was the largest, and the number of clustering at this time was used as the optimal estimated value, which was used as the modal inputs number of FMD calculation model to realize SCBSS of rolling bearing. Simulation and experimental analysis were carried out to verify the effectiveness of the proposed method, and its superiority was also verified through comparison.

Speaker diarization is the process of partitioning an audio recording into segments corresponding to individual speakers. In this paper, we present a robust speaker diarization system and describe its architecture. We focus on discussing the key components necessary for building a strong diarization system, such as voice activity detection (VAD), speaker embedding, and clustering. Our system emerged as the winner in the Voxceleb Speaker Recognition Challenge (VoxSRC) 2023, a widely recognized competition for evaluating speaker diarization systems.

While there has been substantial amount of work in speaker diarization recently, there are few efforts in jointly employing lexical and acoustic information for speaker segmentation. Towards that, we investigate a speaker diarization system using a sequence-to-sequence neural network trained on both lexical and acoustic features. We also propose a loss function that allows for selecting not only the speaker change points but also the best speaker at any time by allowing for different speaker groupings. We incorporate Mel Frequency Cepstral Coefficients (MFCC) as an acoustic feature alongside lexical information that are obtained from conversations from the Fisher dataset. Thus, we show that acoustics provide complementary information to the lexical modality. The experimental results show that sequence-to-sequence system trained on both word sequences and MFCC can improve on speaker diarization result compared to the system that only relies on lexical modality or the baseline MFCC-based system. In addition, we test the performance of our proposed method with Automatic Speech Recognition (ASR) transcripts. While the performance on ASR transcripts drops, the Diarization Error Rate (DER) of our proposed method still outperforms the traditional method based on Bayesian Information Criterion (BIC).

In the recent years, extensive researches have been performed on various possible implementations of speaker diarization systems. These systems require efficient clustering algorithms in order to improve their performances in real-time processing. Teaching–learning-based optimization (TLBO) is such clustering algorithm which can be used to resolve the problem to the optimum clustering in a reasonable time. In this paper, a real-time implementation of speaker diarization (SD) system on raspberry pi 3 (RPi 3) using TLBO technique as classifier has been performed. This system has been evaluated on broadcasting radio dataset (NDTV), and the experimental tests have shown that this technique has succeeded to achieve acceptable performances in terms of diarization error rate (DER = 21.90% and 35% in single- and cross-show diarization, respectively), accuracy (87.30%), and real-time factor (RTF = 2.40). Also, we have tested TLBO technique on a 2.4 GHz Intel Core i5 processor using REPERE corpus. Thus, ameliorated results have been obtained in terms of execution time (xRT) and DER in both tasks of single- and cross-show speaker diarization (0.08 and 0.095, and 18.50% and 26.30%, respectively).

Speaker diarization aims to determine ‘who spoke when’ in a given audio stream. Different applications, such as document structuring or information retrieval have led to the exploration of speaker diarization in many different domains, from broadcast news to lectures, phone conversations and meetings. Almost all current diarization systems are offline and ill-suited to the growing need for online or real-time diarization, stemming from the increasing popularity of powerful, mobile smart devices. While a small number of such systems have been reported, truly online diarization systems for challenging and highly spontaneous meeting data are lacking. This paper reports our work to develop an adaptive and online diarization system using the NIST Rich Transcription meetings corpora. While not dissimilar to those previously reported for less challenging domains, high diarization error rates illustrate the challenge ahead and lead to some ideas to improve performance through future research.

This paper addresses the problem of speaker diarization in the specific context of meeting room recordings. Some new enhancements to the E-HMM-based speaker diarization system are reported. These involve a different approach to speaker modelling utilising EM/ML-based training rather than MAP adaptation as in our previous work. Using the new system we investigate the effects of speech activity detection through speaker diarization experiments conducted on 23 meetings extracted from the NIST/RT evaluation campaign datasets. We propose a new approach, which assigns confidence values according to the type of information carried by the signal and incorporates these values directly into the speaker diarization system. Experimental results show that, perhaps surprisingly, the non-speech segments do not systematically affect the robustness of the speaker diarization system, and more precisely the speaker model training process.

It is difficult to distinguish the mixed signal received by the single channel receiver from single signal simply based on the waveforms and the spectrum features. In this paper, the expression on the observed and the valid signal-to-noise ratio (SNR) of single-channel mixed signal is derived and compared with that of single signal(just one signal component). We find that the observed and the valid SNR of single signal are identical, but for the mixed signal, the observed SNR is significantly greater than valid SNR. An appropriate threshold is set under the condition of minimum the single signal leaking-detecting probability to obtain best detection performance. Numerical results show that for both symmetric and asymmetric mixed signal, the proposed algorithm performs well without timing recovery, thus making proposed algorithm more practical.

Speaker diarization is the task of partitioning an audio stream into homogeneous segments according to speaker identity. Today state-of-the-art speaker diarization systems have achieved very competitive performance. However, any small improvement in Diarization Error Rate (DER) is usually subject to very large processing times (real time factor above one), which makes systems not suitable for some time-critical, real-life applications. Recently, a novel fast speaker diarization technique based on speaker modeling using binary keys was presented. The proposed technique speeds up the process up to ten times faster than real-time with little increase of DER. Although the approach shows great potential, the presented results are still preliminary. The goal of this paper is to further investigate this technique, in order to move towards a complete binary-key based system for the speaker diarization task. Preliminary experiments in Speech Activity Detection (SAD) based on binary keys show the feasibility of the binary key modeling approach for this task. Furthermore, the system has been tested on two different kinds of test data: meeting audio recordings and TV shows. The experiments carried out on NIST RT05 and REPERE databases show promising results and indicate that there is still room for further improvement.

Based on LabVIEW, this paper presents a design scheme of data acquisition, which it proposes using computer and ordinary data collection card as the main hardware and LabVIEW as software development platform, thus constructs the virtual instrument system for signal acquisition. The system can realize synchronous data acquisition of single channel or multi-channel signal, as well as real-time display and preservation. The practice shows that this scheme can not only effectively play performance of ordinary data acquisition card, but also reduce greatly program development time and beautify interface through LabVIEW which has powerfully visual human-machine interface editor and graphical programming function.

The ability to transmit signals with high information spectral density (ISD) using low-complexity and cost-effective transceivers is essential for short- and medium-haul optical communication systems. Consequently, spectrally efficient direct detection transceiver-based solutions are attractive for such applications. In this paper, we experimentally demonstrate the wavelength-division multiplexed (WDM) transmission of 7×12 GHz-spaced dispersion pre-compensated Nyquist pulse-shaped 16-QAM subcarrier modulated channels operating at a net bit rate of 24 Gb/s per channel, and achieving a net optical ISD of 2.0 b/s/Hz. The direct detection receiver used in our experiment consisted of a single-ended photodiode and a single analog-to-digital converter. The carrier-to-signal power ratio at different values of optical signal-to-noise ratio was optimized to maximize the receiver sensitivity performance. The transmission experiments were carried out using a recirculating fiber loop with uncompensated standard single-mode fiber and EDFA-only amplification. The maximum achieved transmission distances for single channel and WDM signals were 727 and 323 km below the bit-error ratio of 3.8 × 10-3, respectively. To the best of our knowledge, this is the highest achieved ISD for WDM transmission in direct detection links over such distances.