Search

The I4U consortium was established to facilitate a joint entry to NIST speaker recognition evaluations (SRE). The latest edition of such joint submission was in SRE 2018, in which the I4U submission was among the best-performing systems. SRE'18 also marks the 10-year anniversary of I4U consortium into NIST SRE series of evaluation. The primary objective of the current paper is to summarize the results and lessons learned based on the twelve subsystems and their fusion submitted to SRE'18. It is also our intention to present a shared view on the advancements, progresses, and major paradigm shifts that we have witnessed as an SRE participant in the past decade from SRE'08 to SRE'18. In this regard, we have seen, among others , a paradigm shift from supervector representation to deep speaker embedding, and a switch of research challenge from channel compensation to domain adaptation.

In this study, we present systems submitted by the Center for Robust Speech Systems (CRSS) from UTDallas to NIST SRE 2018 (SRE18). Three alternative front-end speaker embedding frameworks are investigated, that includes: (i) i-vector, (ii) x-vector, (iii) and a modified triplet speaker embedding system (t-vector). Similar to the previous SRE, language mismatch between training and enrollment/test data, the so-called domain mismatch, remains as a major challenge in this evaluation. In addition, SRE18 also introduces a small portion of audio from an unstructured video corpus in which speaker detection/diarization is supposedly needed to be effectively integrated into speaker recognition for system robustness. In our system development, we focused on: (i) building novel deep neural network based speaker discriminative embedding systems as utterance level feature representations, (ii) exploring alternative dimension reduction methods, back-end classifiers, score normalization techniques which can incorporate unlabeled in-domain data for domain adaptation, (iii) finding an improved data set configurations for the speaker embedding network, LDA/PLDA, and score calibration training (v) and finally, investigating effective score calibration and fusion strategies. The final resulting systems are shown to be both complementary and effective in achieving overall improved speaker recognition performance.

Speech separation has been studied widely for single-channel close-talk microphone recordings over the past few years; developed solutions are mostly in frequency-domain. Recently, a raw audio waveform separation network (TasNet) is introduced for single-channel data, with achieving high Si-SNR (scale-invariant source-to-noise ratio) and SDR (source-to-distortion ratio) comparing against the state-of-the-art solution in frequency-domain. In this study, we incorporate effective components of the TasNet into a frequency-domain separation method. We compare both for alternative scenarios. We introduce a solution for directly optimizing the separation criterion in frequency-domain networks. In addition to speech separation objective and subjective measurements, we evaluate the separation performance on a speech recognition task as well. We study the speech separation problem for far-field data (more similar to naturalistic audio streams) and develop multi-channel solutions for both frequency and time-domain separators with utilizing spectral, spatial and speaker location information. For our experiments, we simulated multi-channel spatialized reverberate WSJ0-2mix dataset. Our experimental results show that spectrogram separation can achieve competitive performance with better network design. Multi-channel framework as well is shown to improve the single-channel performance relatively up to +35.5% and +46% in terms of WER and SDR, respectively., Comment: INTERSPEECH 2019

This study presents systems submitted by the University of Texas at Dallas, Center for Robust Speech Systems (UTD-CRSS) to the MGB-3 Arabic Dialect Identification (ADI) subtask. This task is defined to discriminate between five dialects of Arabic, including Egyptian, Gulf, Levantine, North African, and Modern Standard Arabic. We develop multiple single systems with different front-end representations and back-end classifiers. At the front-end level, feature extraction methods such as Mel-frequency cepstral coefficients (MFCCs) and two types of bottleneck features (BNF) are studied for an i-Vector framework. As for the back-end level, Gaussian back-end (GB), and Generative Adversarial Networks (GANs) classifiers are applied alternately. The best submission (contrastive) is achieved for the ADI subtask with an accuracy of 76.94% by augmenting the randomly chosen part of the development dataset. Further, with a post evaluation correction in the submitted system, final accuracy is increased to 79.76%, which represents the best performance achieved so far for the challenge on the test dataset.

This article proposes a method to improve the performance of deterministic plus stochastic model (DSM-) based feature extraction by integrating the contextual information. One precious advantage of speech synthesis over speech recognition is that in both training and testing phases of synthesis, contextual information is available. However, similar to recognition, this invaluable knowledge has been forgotten during acoustic feature extraction of speech synthesis. DSM expresses the residual of Mel-cepstral analysis through a summation of two components, namely deterministic and stochastic. This study proposes to model the deterministic component through a novel context-dependent principal component analysis (CD-PCA), and the stochastic component through the conventional high-pass filtered noise. Furthermore, due to the high dependency of the proposed feature extraction on state boundaries, the feature analysis and HMM-based modeling are performed in an iterative manner. Subjective evaluations conducted on a Persian speech database confirm the effectiveness of the proposed synthesis system.

Hidden Markov Model (HMM)-based synthesis (HTS) has recently been confirmed to be the most effective method in generating natural speech. However, it lacks adequate context generalization when the training data is limited. As a solution, current study provides a new context-dependent speech modeling framework based on the Gaussian Conditional Random Field (GCRF) theory. By applying this model, an innovative speech synthesis system has been developed which can be viewed as an extension of Context-Dependent Hidden Semi Markov Model (CD-HSMM). A novel Viterbi decoder along with a stochastic gradient ascent algorithm was applied to train model parameters. Also, a fast and efficient parameter generation algorithm was derived for the synthesis part. Experimental results using objective and subjective criteria have shown that the proposed system outperforms HSMM substantially in limited speech databases. Moreover, Mel-cepstral distance of the spectral parameters has been reduced considerably for any size of training database.

Speech synthesis systems provided for the Persian language so far need various large-scale speech corpora to synthesize several target speakers' voice. Accordingly, synthesizing speech with a small amount of data seems to be essential in Persian. Taking advantage of a speaker adaptation in the speech synthesis systems makes it possible to generate speech with remarkable quality when the data of the speaker are limited. Here we conducted this method for the first time in Persian. This paper describes speaker adaptation based on Hidden Markov Models (HMMs) in Persian speech synthesis system for FARsi Speech DATabase (FARSDAT). In this regard, we prepared the whole FARSDAT, then for synthesizing speech with arbitrary speaker characteristics, we trained the average voice units; afterward, the adapted model was obtained by transforming the average voice model. We demonstrate that a few speech data of a target speaker are sufficient to obtain high quality synthetic speech, and we set out synthetic speech which has been generated from adapted models by using only 88 utterances is very close to that from speaker dependent models trained using 355 utterances.