Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Voice separation method based on parameterized multi-phase gammatone filter bank

A filter bank, speech separation technology, used in speech analysis, instrumentation, etc., to solve problems such as suboptimal performance

Active Publication Date: 2021-07-06
NORTHWESTERN POLYTECHNICAL UNIV +1
View PDF2 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

However, this method does not achieve the best performance in speech separation, and there is room for further improvement

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Voice separation method based on parameterized multi-phase gammatone filter bank
  • Voice separation method based on parameterized multi-phase gammatone filter bank
  • Voice separation method based on parameterized multi-phase gammatone filter bank

Examples

Experimental program
Comparison scheme
Effect test

specific Embodiment

[0064] (1) Experimental settings:

[0065] The Conv-Tasnet network is trained for 200 epochs on 4-second long segments. The optimizer adopts Adam optimizer, and the initial learning rate is 0.001. If the performance does not improve for 5 consecutive epochs on the validation set, the learning rate is halved. Also, when the performance on the validation set has not improved in the past 10 epochs, the network training will be stopped. The hyperparameter setting of the network follows the network hyperparameters in Conv-Tasnet, where the number of filters N is 512. The mask functions of Temporal Convolutional Networks (TCN) are set as sigmoid function and rectified linear unit (ReLU) respectively. For ParaMPGTF, the order n is set to 2 and the magnitude α is set to 1. will c 1 and c 2 The initial value of is set to its empirical value, namely c 1 =24.7,c 2 =9.265. SI-SNR is used as the evaluation index. The reported results are the average results of 3000 sentences of t...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses a voice separation method based on a parameterized multi-phase gammatone filter bank. The method comprises the following steps of: firstly, constructing the parameterized multi-phase gammatone filter bank on the basis of a gammatone filter, then replacing an encoder of a Conv-Tasnet network with the parameterized multi-phase gammatone filter bank, keeping the encoder unchanged or adopting inverse transformation of the parameterized multi-phase gammatone filter bank to form a new Conv-Tasnet network, and training the new Conv-Tasnet network to obtain a final voice separation network. According to the method provided by the invention, under the condition that the decoder has learnable features, competitive performance is obtained; and under the condition that a decoder is inverse transformation of an encoder, the feature is superior to the features of manual design such as STFT and MPGTF.

Description

technical field [0001] The invention belongs to the technical field of voice recognition, and in particular relates to a voice separation method. Background technique [0002] The purpose of speech separation is to separate the mixed speech of multiple sound sources into its corresponding components. In recent years, various methods such as deep clustering, permutation invariant training, and deep attractor network have been proposed for the problem of speech separation. However, in these methods, the widely used acoustic feature is the amplitude spectrum of the short-time Fourier transform (short–time Fourier transform, STFT). This leads to the use of the noisy phase spectrum in the process of restoring the time domain signal from the separated magnitude spectrum, resulting in suboptimal performance. [0003] To overcome this shortcoming, learnable features transformed from time domain to time-frequency domain learned by the network become a new trend. The representative...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G10L21/0272
CPCG10L21/0272Y02E40/40
Inventor 张晓雷朱文博王逸平
Owner NORTHWESTERN POLYTECHNICAL UNIV
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com