Search

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.

The recently-proposed deep clustering-based algorithms represent a fundamental advance towards the single-channel multi-speaker speech sep- aration problem. These methods use an ideal binary mask to construct the objective function and K-means clustering method to estimate the ideal bina- ry mask. However, when sources belong to the same class or the number of sources is large, the assumption that one time-frequency unit of the mixture is dominated by only one source becomes weak, and the IBM-based separation causes spectral holes or aliasing. Instead, in our work, the quantized ideal ratio mask was proposed, the ideal ratio mask is quantized to have the output of the neural network with a limited number of possible values. Then the quan- tized ideal ratio mask is used to construct the objective function for the case of multi-source domination, to improve network performance. Furthermore, a network framework that combines a residual network, a recurring network, and a fully connected network was used for exploiting correlation information of frequency in our work. We evaluated our system on TIMIT dataset and show 1.6 dB SDR improvement over the previous state-of-the-art methods.

Object-based audio content is becoming the main form of audio content, because it is more interactive and flexible than traditional channel-based audio content. The Spatial Audio Object Coding (SAOC) method is proposed to encode multiple audio objects at low bitrate. However, SAOC extracts only a few parameters for each frame signal, which leads to low parameter frequency resolution. So the decoded signals have serious aliasing distortion which will destroy the sound quality. In this paper, we present a novel audio object coding method. We are the first to analyze how the signal distortion varies with parameter frequency resolution, and determine the optimal resolution to reduce aliasing distortion. In addition, we also achieve low coding bitrate by the dimensional reduction algorithm. Both the objective and subjective experiments confirm that the proposed method can provide higher sound quality of output signals than the state-of-the-art methods at equivalent bitrate.