文档智能

最新阿里云云产品活动优惠券领取，阿里云文档智能基于多年技术积累打造的多模态文档识别与理解引擎，为用户提供各类文档文字提取和文档处理，支持通用场景、行业场景和自定义场景下的多样化文档处理需求。

To jointly optimize the waveform-domain adversarial loss, we employ multi-period discriminator (MPD)[20-21] and multi-scale discriminator (MSD)[20-21] to identify speech signal from two different perspectives. The MSD method is derived from the MelGAN vocoder. Through the average pooling operation, the length of the speech sequence is halved successively. Then the convolution operation is performed on the speech signals of different scales. Finally, it is flattened and output. The MPD method folds the single-channel audio sequence into a two-channel audio with different fixed-lengths called period, and then apply 2-D convolution on the folded data. The disadvantage of this approach is the folded data on each channel is mixed with artifacts of different frequencies. In edy order to make up for this defect, we proposed multi-length discriminator (MLD), to improve ability of discriminating synthetic or real audio as much as possible. Firstly, the single-channel audio is folded into multi-channel audio by wavelet transform[20]. Then apply 1-D dilated convolution as in [22]. In this way, each channel in the folded data contains few or no artifacts of other frequencies, ensuring the stability and accuracy of the discrimination.

The generator of PLCNet is a symmetric encoder-decoder structure with skip connections and residual units. The encoder and decoder each have 4 sub-modules, each sub-module of the encoder consists of 3 residual units and a down-sampling module and each sub-module of the decoder consists of an up-sampling module and 3 residual units. The residual unit alternately uses 1-D dilated convolution with kernel size of 7and 1-D convolution with kernel size of 1. The dilation rate is gradually increased using (1,3, 9). The input is first transformed by 1-D convolution with kernel size of 7, then the encoder maps the 16khz waveform to the 50hz representation through down-sampling block of (2,4,5,8) in the form of a stride convolution. The decoder uses the transposed convolution method to up-sampling in reverse order, restores the features to the same dimension as the speech. The number of channels is doubled when down-sampling and halved when up-sampling. The middle bottleneck layer acts as a bridge between encoder and decoder and consists of 3 1-D convolutions with kernel size of7. A skip-connection is used between the corresponding layers of the encoder and decoder to allow information such as phase or alignment to pass through. We use the ELU activation function [19] and weight normalization in the generator to guarantee the stability of adversarial training. Finally, the output of the decoder is a mono signal, with tanh limiting the output range to [-1,1]. To be able to process real-time audio streams on low-power mobile devices, all our convolutions are causal.