Automatic discrimination between front and back ensemble locations in HRTF-convolved binaural recordings of music

One of the greatest challenges in the development of binaural machine audition systems is the disambiguation between front and back audio sources, particularly in complex spatial audio scenes. The goal of this work was to develop a method for discriminating between front and back located ensembles in binaural recordings of music. To this end, 22, 496 binaural excerpts, representing either front or back located ensembles, were synthesized by convolving multi-track music recordings with 74 sets of head-related transfer functions (HRTF). The discrimination method was developed based on the traditional approach, involving hand-engineering of features, as well as using a deep learning technique incorporating the convolutional neural network (CNN). According to the results obtained under HRTF-dependent test conditions, CNN showed a very high discrimination accuracy (99.4%), slightly outperforming the traditional method. However, under the HRTF-independent test scenario, CNN performed worse than the traditional algorithm, highlighting the importance of testing the algorithms under HRTF-independent conditions and indicating that the traditional method might be more generalizable than CNN. A minimum of 20 HRTFs are required to achieve a satisfactory generalization performance for the traditional algorithm and 30 HRTFs for CNN. The minimum duration of audio excerpts required by both the traditional and CNN-based methods was assessed as 3 s. Feature importance analysis, based on a gradient attribution mapping technique, revealed that for both the traditional and the deep learning methods, a frequency band between 5 and 6 kHz is particularly important in terms of the discrimination between front and back ensemble locations. Linear-frequency cepstral coefficients, interaural level differences, and audio bandwidth were identified as the key descriptors facilitating the discrimination process using the traditional approach.

Improving low-resource Tibetan end-to-end ASR by multilingual and multilevel unit modeling

Conventional automatic speech recognition (ASR) and emerging end-to-end (E2E) speech recognition have achieved promising results after being provided with sufficient resources. However, for low-resource language, the current ASR is still challenging. The Lhasa dialect is the most widespread Tibetan dialect and has a wealth of speakers and transcriptions. Hence, it is meaningful to apply the ASR technique to the Lhasa dialect for historical heritage protection and cultural exchange. Previous work on Tibetan speech recognition focused on selecting phone-level acoustic modeling units and incorporating tonal information but underestimated the influence of limited data. The purpose of this paper is to improve the speech recognition performance of the low-resource Lhasa dialect by adopting multilingual speech recognition technology on the E2E structure based on the transfer learning framework. Using transfer learning, we first establish a monolingual E2E ASR system for the Lhasa dialect with different source languages to initialize the ASR model to compare the positive effects of source languages on the Tibetan ASR model. We further propose a multilingual E2E ASR system by utilizing initialization strategies with different source languages and multilevel units, which is proposed for the first time. Our experiments show that the performance of the proposed method-based ASR system exceeds that of the E2E baseline ASR system. Our proposed method effectively models the low-resource Lhasa dialect and achieves a relative 14.2% performance improvement in character error rate (CER) compared to DNN-HMM systems. Moreover, from the best monolingual E2E model to the best multilingual E2E model of the Lhasa dialect, the system’s performance increased by 8.4% in CER.

Auxiliary function-based algorithm for blind extraction of a moving speaker

In this paper, we propose a novel algorithm for blind source extraction (BSE) of a moving acoustic source recorded by multiple microphones. The algorithm is based on independent vector extraction (IVE) where the contrast function is optimized using the auxiliary function-based technique and where the recently proposed constant separating vector (CSV) mixing model is assumed. CSV allows for movements of the extracted source within the analyzed batch of recordings. We provide a practical explanation of how the CSV model works when extracting a moving acoustic source. Then, the proposed algorithm is experimentally verified on the task of blind extraction of a moving speaker. The algorithm is compared with state-of-the-art blind methods and with an adaptive BSE algorithm which processes data in a sequential manner. The results confirm that the proposed algorithm can extract the moving speaker better than the BSE methods based on the conventional mixing model and that it achieves improved extraction accuracy than the adaptive method.

Anchor voiceprint recognition in live streaming via RawNet-SA and gated recurrent unit

With the sharp booming of online live streaming platforms, some anchors seek profits and accumulate popularity by mixing inappropriate content into live programs. After being blacklisted, these anchors even forged their identities to change the platform to continue live, causing great harm to the network environment. Therefore, we propose an anchor voiceprint recognition in live streaming via RawNet-SA and gated recurrent unit (GRU) for anchor identification of live platform. First, the speech of the anchor is extracted from the live streaming by using voice activation detection (VAD) and speech separation. Then, the feature sequence of anchor voiceprint is generated from the speech waveform with the self-attention network RawNet-SA. Finally, the feature sequence of anchor voiceprint is aggregated by GRU to transform into a deep voiceprint feature vector for anchor recognition. Experiments are conducted on the VoxCeleb, CN-Celeb, and MUSAN dataset, and the competitive results demonstrate that our method can effectively recognize the anchor voiceprint in video streaming.

A recursive expectation-maximization algorithm for speaker tracking and separation

The problem of blind and online speaker localization and separation using multiple microphones is addressed based on the recursive expectation-maximization (REM) procedure. A two-stage REM-based algorithm is proposed: (1) multi-speaker direction of arrival (DOA) estimation and (2) multi-speaker relative transfer function (RTF) estimation. The DOA estimation task uses only the time frequency (TF) bins dominated by a single speaker while the entire frequency range is not required to accomplish this task. In contrast, the RTF estimation task requires the entire frequency range in order to estimate the RTF for each frequency bin. Accordingly, a different statistical model is used for the two tasks. The first REM model is applied under the assumption that the speech signal is sparse in the TF domain, and utilizes a mixture of Gaussians (MoG) model to identify the TF bins associated with a single dominant speaker. The corresponding DOAs are estimated using these bins. The second REM model is applied under the assumption that the speakers are concurrently active in all TF bins and consequently applies a multichannel Wiener filter (MCWF) to separate the speakers. As a result of the assumption of the concurrent speakers, a more precise TF map of the speakers’ activity is obtained. The RTFs are estimated using the outputs of the MCWF-beamformer (BF), which are constructed using the DOAs obtained in the previous stage. Next, using the linearly constrained minimum variance (LCMV)-BF that utilizes the estimated RTFs, the speech signals are separated. The algorithm is evaluated using real-life scenarios of two speakers. Evaluation of the mean absolute error (MAE) of the estimated DOAs and the separation capabilities, demonstrates significant improvement w.r.t. a baseline DOA estimation and speaker separation algorithm.

Text-to-speech system for low-resource language using cross-lingual transfer learning and data augmentation

Deep learning techniques are currently being applied in automated text-to-speech (TTS) systems, resulting in significant improvements in performance. However, these methods require large amounts of text-speech paired data for model training, and collecting this data is costly. Therefore, in this paper, we propose a single-speaker TTS system containing both a spectrogram prediction network and a neural vocoder for the target language, using only 30 min of target language text-speech paired data for training. We evaluate three approaches for training the spectrogram prediction models of our TTS system, which produce mel-spectrograms from the input phoneme sequence: (1) cross-lingual transfer learning, (2) data augmentation, and (3) a combination of the previous two methods. In the cross-lingual transfer learning method, we used two high-resource language datasets, English (24 h) and Japanese (10 h). We also used 30 min of target language data for training in all three approaches, and for generating the augmented data used for training in methods 2 and 3. We found that using both cross-lingual transfer learning and augmented data during training resulted in the most natural synthesized target speech output. We also compare single-speaker and multi-speaker training methods, using sequential and simultaneous training, respectively. The multi-speaker models were found to be more effective for constructing a single-speaker, low-resource TTS model. In addition, we trained two Parallel WaveGAN (PWG) neural vocoders, one using 13 h of our augmented data with 30 min of target language data and one using the entire 12 h of the original target language dataset. Our subjective AB preference test indicated that the neural vocoder trained with augmented data achieved almost the same perceived speech quality as the vocoder trained with the entire target language dataset. Overall, we found that our proposed TTS system consisting of a spectrogram prediction network and a PWG neural vocoder was able to achieve reasonable performance using only 30 min of target language training data. We also found that by using 3 h of target language data, for training the model and for generating augmented data, our proposed TTS model was able to achieve performance very similar to that of the baseline model, which was trained with 12 h of target language data.

Spherical harmonic covariance and magnitude function encodings for beamformer design

Microphone and speaker array designs have increasingly diverged from simple topologies due to diversity of physical host geometries and use cases. Effective beamformer design must now account for variation in the array’s acoustic radiation pattern, spatial distribution of target and noise sources, and intended beampattern directivity. Relevant tasks such as representing complex pressure fields, specifying spatial priors, and composing beampatterns can be efficiently synthesized using spherical harmonic (SH) basis functions. This paper extends the expansion of common stationary covariance functions onto the SHs and proposes models for encoding magnitude functions on a sphere. Conventional beamformer designs are reformulated in terms of magnitude density functions and beampatterns along SH bases. Applications to speaker far-field response fitting, cross-talk cancelation design, and microphone beampattern fitting are presented.

Unsupervised domain adaptation for lip reading based on cross-modal knowledge distillation

We present an unsupervised domain adaptation (UDA) method for a lip-reading model that is an image-based speech recognition model. Most of conventional UDA methods cannot be applied when the adaptation data consists of an unknown class, such as out-of-vocabulary words. In this paper, we propose a cross-modal knowledge distillation (KD)-based domain adaptation method, where we use the intermediate layer output in the audio-based speech recognition model as a teacher for the unlabeled adaptation data. Because the audio signal contains more information for recognizing speech than lip images, the knowledge of the audio-based model can be used as a powerful teacher in cases where the unlabeled adaptation data consists of audio-visual parallel data. In addition, because the proposed intermediate-layer-based KD can express the teacher as the sub-class (sub-word)-level representation, this method allows us to use the data of unknown classes for the adaptation. Through experiments on an image-based word recognition task, we demonstrate that the proposed approach can not only improve the UDA performance but can also use the unknown-class adaptation data.

U2-VC: one-shot voice conversion using two-level nested U-structure

Voice conversion is to transform a source speaker to the target one, while keeping the linguistic content unchanged. Recently, one-shot voice conversion gradually becomes a hot topic for its potentially wide range of applications, where it has the capability to convert the voice from any source speaker to any other target speaker even when both the source speaker and the target speaker are unseen during training. Although a great progress has been made in one-shot voice conversion, the naturalness of the converted speech remains a challenging problem. To further improve the naturalness of the converted speech, this paper proposes a two-level nested U-structure (U2-Net) voice conversion algorithm called U2-VC. The U2-Net can extract both local feature and multi-scale feature of log-mel spectrogram, which can help to learn the time-frequency structures of the source speech and the target speech. Moreover, we adopt sandwich adaptive instance normalization (SaAdaIN) in decoder for speaker identity transformation to retain more content information of the source speech while maintaining the speaker similarity between the converted speech and the target speech. Experiments on VCTK dataset show that U2-VC outperforms many SOTA approaches including AGAIN-VC and AdaIN-VC in terms of both objective and subjective measurements.

dEchorate: a calibrated room impulse response dataset for echo-aware signal processing

This paper presents a new dataset of measured multichannel room impulse responses (RIRs) named dEchorate. It includes annotations of early echo timings and 3D positions of microphones, real sources, and image sources under different wall configurations in a cuboid room. These data provide a tool for benchmarking recent methods in echo-aware speech enhancement, room geometry estimation, RIR estimation, acoustic echo retrieval, microphone calibration, echo labeling, and reflector position estimation. The dataset is provided with software utilities to easily access, manipulate, and visualize the data as well as baseline methods for echo-related tasks.