Electrical Engineering And Systems Science

We describe the speech activity detection (SAD), speaker diarization (SD), and automatic speech recognition (ASR) experiments conducted by the Behavox team for the Interspeech 2020 Fearless Steps Challenge (FSC-2). A relatively small amount of labeled data, a large variety of speakers and channel distortions, specific lexicon and speaking style resulted in high error rates on the systems which involved this data. In addition to approximately 36 hours of annotated NASA mission recordings, the organizers provided a much larger but unlabeled 19k hour Apollo-11 corpus that we also explore for semi-supervised training of ASR acoustic and language models, observing more than 17% relative word error rate improvement compared to training on the FSC-2 data only. We also compare several SAD and SD systems to approach the most difficult tracks of the challenge (track 1 for diarization and ASR), where long 30-minute audio recordings are provided for evaluation without segmentation or speaker information. For all systems, we report substantial performance improvements compared to the FSC-2 baseline systems, and achieved a first-place ranking for SD and ASR and fourth-place for SAD in the challenge.

In addition to being extremely non-linear, modern problems require millions if not billions of parameters to solve or at least to get a good approximation of the solution, and neural networks are known to assimilate that complexity by deepening and widening their topology in order to increase the level of non-linearity needed for a better approximation. However, compact topologies are always preferred to deeper ones as they offer the advantage of using less computational units and less parameters. This compacity comes at the price of reduced non-linearity and thus, of limited solution search space. We propose the 1-Dimensional Polynomial Neural Network (1DPNN) model that uses automatic polynomial kernel estimation for 1-Dimensional Convolutional Neural Networks (1DCNNs) and that introduces a high degree of non-linearity from the first layer which can compensate the need for deep and/or wide topologies. We show that this non-linearity introduces more computational complexity but enables the model to yield better results than a regular 1DCNN that has the same number of training parameters on various classification and regression problems related to audio signals. The experiments were conducted on three publicly available datasets and demonstrate that the proposed model can achieve a much faster convergence than a 1DCNN on the tackled regression problems.

A microphone array can provide a mobile robot with the capability of localizing, tracking and separating distant sound sources in 2D, i.e., estimating their relative elevation and azimuth. To combine acoustic data with visual information in real world settings, spatial correlation must be established. The approach explored in this paper consists of having two robots, each equipped with a microphone array, localizing themselves in a shared reference map using SLAM. Based on their locations, data from the microphone arrays are used to triangulate in 3D the location of a sound source in relation to the same map. This strategy results in a novel cooperative sound mapping approach using mobile microphone arrays. Trials are conducted using two mobile robots localizing a static or a moving sound source to examine in which conditions this is possible. Results suggest that errors under 0.3 m are observed when the relative angle between the two robots are above 30 degrees for a static sound source, while errors under 0.3 m for angles between 40 degrees and 140 degrees are observed with a moving sound source.

Modern automatic speaker verification relies largely on deep neural networks (DNNs) trained on mel-frequency cepstral coefficient (MFCC) features. While there are alternative feature extraction methods based on phase, prosody and long-term temporal operations, they have not been extensively studied with DNN-based methods. We aim to fill this gap by providing extensive re-assessment of 14 feature extractors on VoxCeleb and SITW datasets. Our findings reveal that features equipped with techniques such as spectral centroids, group delay function, and integrated noise suppression provide promising alternatives to MFCCs for deep speaker embeddings extraction. Experimental results demonstrate up to 16.3\% (VoxCeleb) and 25.1\% (SITW) relative decrease in equal error rate (EER) to the baseline.

We develop a virtual analog model of the Klon Centaur guitar pedal circuit, comparing various circuit modelling techniques. The techniques analyzed include traditional modelling techniques such as nodal analysis and Wave Digital Filters, as well as a machine learning technique using recurrent neural networks. We examine these techniques in the contexts of two use cases: an audio plug-in designed to be run on a consumer-grade desktop computer, and a guitar pedal-style effect running on an embedded device. Finally, we discuss the advantages and disdvantages of each technique for modelling different circuits, and targeting different platforms.

A DJ mix is a sequence of music tracks concatenated seamlessly, typically rendered for audiences in a live setting by a DJ on stage. As a DJ mix is produced in a studio or the live version is recorded for music streaming services, computational methods to analyze DJ mixes, for example, extracting track information or understanding DJ techniques, have drawn research interests. Many of previous works are, however, limited to identifying individual tracks in a mix or segmenting it, and the sizes of the datasets are usually small. In this paper, we provide an in-depth analysis of DJ music by aligning a mix to its original music tracks. We set up the subsequence alignment such that the audio features are less sensitive to the tempo or key change of the original track in a mix. This approach provides temporally tight mix-to-track matching from which we can obtain cue-points, transition length, mix segmentation, and musical changes in DJ performance. Using 1,557 mixes from 1001Tracklists including 13,728 tracks and 20,765 transitions, we conduct the proposed analysis and show a wide range of statistics, which may elucidate the creative process of DJ music making.

We present an open-source speech corpus for the Kazakh language. The Kazakh speech corpus (KSC) contains around 332 hours of transcribed audio comprising over 153,000 utterances spoken by participants from different regions and age groups, as well as both genders. It was carefully inspected by native Kazakh speakers to ensure high quality. The KSC is the largest publicly available database developed to advance various Kazakh speech and language processing applications. In this paper, we first describe the data collection and preprocessing procedures followed by a description of the database specifications. We also share our experience and challenges faced during the database construction, which might benefit other researchers planning to build a speech corpus for a low-resource language. To demonstrate the reliability of the database, we performed preliminary speech recognition experiments. The experimental results imply that the quality of audio and transcripts is promising (2.8% character error rate and 8.7% word error rate on the test set). To enable experiment reproducibility and ease the corpus usage, we also released an ESPnet recipe for our speech recognition models.

Objectives: Evaluation of hypernasality requires extensive perceptual training by clinicians and extending this training on a large scale internationally is untenable; this compounds the health disparities that already exist among children with cleft. In this work, we present the objective hypernasality measure (OHM), a speech analytics algorithm that automatically measures hypernasality in speech, and validate it relative to a group of trained clinicians. Methods: We trained a deep neural network (DNN) on approximately 100 hours of a publicly-available healthy speech corpus to detect the presence of nasal acoustic cues generated through the production of nasal consonants and nasalized phonemes in speech. Importantly, this model does not require any clinical data for training. The posterior probabilities of the deep learning model were aggregated at the sentence and speaker-levels to compute the OHM. Results: The results showed that the OHM was significantly correlated with the perceptual hypernasality ratings in the Americleft database ( r=0.797, ~p < 0.001), and with the New Mexico Cleft Palate Center (NMCPC) database (r=0.713,p<$0.001). In addition, we evaluated the relationship between the OHM and articulation errors; the sensitivity of the OHM in detecting the presence of very mild hypernasality; and establishing the internal reliability of the metric. Further, the performance of OHM was compared with a DNN regression algorithm directly trained on the hypernasal speech samples. Significance: The results indicate that the OHM is able to rate the severity of hypernasality on par with Americleft-trained clinicians on this dataset.

Unison singing is the name given to an ensemble of singers simultaneously singing the same melody and lyrics. While each individual singer in a unison sings the same principle melody, there are slight timing and pitch deviations between the singers, which, along with the ensemble of timbres, give the listener a perceived sense of "unison". In this paper, we present a study of unison singing in the context of choirs; utilising some recently proposed deep-learning based methodologies, we analyse the fundamental frequency (F0) distribution of the individual singers in recordings of unison mixtures. Based on the analysis, we propose a system for synthesising a unison signal from an a cappella input and a single voice prototype representative of a unison mixture. We use subjective listening tests to evaluate perceptual factors of our proposed system for synthesis, including quality, adherence to the melody as well the degree of perceived unison.

This paper proposes a generalized framework for domain adaptation of Probabilistic Linear Discriminant Analysis (PLDA) in speaker recognition. It not only includes several existing supervised and unsupervised domain adaptation methods but also makes possible more flexible usage of available data in different domains. In particular, we introduce here the two new techniques described below. (1) Correlation-alignment-based interpolation and (2) covariance regularization. The proposed correlation-alignment-based interpolation method decreases minCprimary up to 30.5% as compared with that from an out-of-domain PLDA model before adaptation, and minCprimary is also 5.5% lower than with a conventional linear interpolation method with optimal interpolation weights. Further, the proposed regularization technique ensures robustness in interpolations w.r.t. varying interpolation weights, which in practice is essential.