The efficient transmission of the audio objects can be achieved by spatial audio object coding (SAOC) method that conveys a mono downmix signal together with side information parameters that enable object reconstruction in the decoder. To allow the transmission of audio objects at low bitrates, we present a new audio coding method with convolutional auto-encoder (CAE) and dense convolutional network (DenseNet) mixture model, optimizing the compression of side information parameters of audio objects. It has two main advantages: 1) Different from the linear transform methods, CAE can dig the nonlinear relationship of side information parameters and can effectively reduce the dimension of side information parameters; 2) DenseNet is adding in the decoder to make full use of low dimensional features of side information parameters, which improves the audio quality at low bitrate. Experiments show that our method outperforms base-line methods permitting bitrates as low as 1 kbps per object.

Speech communications and interactions frequently occur in a variety of environments. Noise in the environment significantly degrades speech intelligibility when speaking and listening. Especially in the listening stage, even if the multimedia terminal outputs clean speech, it is still difficult for listeners to obtain information. Intelligibility enhancement (IENH) of speech is a technique for overcoming the environmental noise in the listening stage. It implements a perceptual enhancement of non-noisy speech. This study focuses on IENH via normal-to-Lombard speech conversion, inspired by a well known acoustic mechanism named the Lombard effect. Our method combines the long short-term memory (LSTM) network and Bayesian Gaussian mixture model (BGMM) to build a conversion architecture. Compared with baselines, it has three main advantages: 1) an LSTM network is used for spectral tilt mapping with fully considering short-term correlations and high-dimensional expression abilities; 2) the aperiodicity (AP) is mapped together with the fundamental frequency ($F_0$) by a BGMM, which considers their relevance constraints and the importance of APs; 3) the gender-dependent mapping is used for$F_0$and APs to consider distribution differences between genders. Experiments indicate that our method gets better performance in both objective and subjective tests.

Speech intelligibility enhancement is a perceptual enhancement technique for clean speech reproduced in noisy environments. It is typically used in the listening stage of multimedia communications. In this study, we enhance speech intelligibility by speaking style conversion (SSC), which is a data-driven approach inspired by a vocal mechanism named Lombard effect. The proposed SSC method combines star generative adversarial network (StarGAN) based mapping and dynamic range compression (DRC). It has two main advantages: 1) different from gender-independent conversion in previous studies, StarGAN can separately learn speech features of different genders to provide a differential conversion among genders with a single model and non-parallel training data; 2) we design a multi-level enhancement strategy with the use of DRC in the StarGAN architecture, which improves the SSC performance in strong noise interference. Experiments show that our method outperforms baseline methods.

Noise in the environment significantly decreases the speech intelligibility of telephone conversations. Despite clean speech output from the device, the listener is still hard to get information. This study focuses on intelligibility enhancement (IENH) of telephone speech in near-end background noise based on normal-to-Lombard speech conversion. The proposed approach uses long short-term memory (LSTM) and Bayesian Gaussian mixture model (BGMM) to build the speech mapping model. Compared with previous studies, we fully consider the short-term correlations of speech and implement feature mappings with higher dimensional features and more types of features. Evaluations indicate that the proposed approach has achieved better results in both objective and subjective evaluation.

In this paper, a new attractor points estimation method for DANet algorithm used in single channel multi-speaker speech separation has been proposed. A prerequisite is that there must be separate segments of each source in the mixture. This condition is met in the actual situation because the source signal is not overlapping at any time. With this prerequisite, an isolated source segments extracted from the mixture is converted to the embedding space. With the embedding of isolated source segments, a more accurately attractor point for each source will be created, due to it does not contain components of other sources. In addition, a gaussian mixture model(GMM) clustering method instead of K-means clustering method were used at run time. The experiment demonstrated that the proposed method gets a better separation performance than state of the art method up to 1.04dB in SDR.