Two-level-subspace partitioning method and device

Abstract

The invention discloses a two-level-subspace partitioning method and device in the technical scheme. The two-level-subspace partitioning method and device are used for achieving the aims that a secondary matching method based on model subspaces is adopted; the processing process includes the two steps that the subspaces are positioned through a coarse screen, and a range of a to-be-detected audiofrequency module space is determined; fine matching, wherein a target model is hit inside the subspaces through a high-accuracy traditional algorithm, and four processes of feature extraction, clustering calculation, subspace partitioning and central point calculation are conducted. The two-level-subspace partitioning method and device has the advantages that achievement is simple, less time consumption is wasted, and for model features with the large quantity, the real-time performance of the device enough meets the business requirements, values of the cluster quantity can be adjusted at random, the subspaces can be effectively partitioned, the values can be evenly partitioned into different subspaces, and meanwhile the model matching range is narrowed; the quantity of models can be effectively reduced through single voice features, and the efficiency is improved.

Description

technical field [0001] The invention relates to a two-level subspace division method and device, belonging to the technical field of computers. Background technique [0002] Currently based on the floating-point comparison algorithm, it takes a lot of time. After the system exceeds 2000 voice models, its real-time performance is not enough to meet business needs. With the evolution of the system and the increase of the model feature library, based on the comparison of large-scale speech model features, it poses challenges to the traditional comparison algorithm, which cannot meet the real-time requirements of the system. Therefore, a two-level comparison idea is proposed, in which the division of subspaces becomes a difficult point in calculation, how to divide the models in the model library into different subspaces. In the existing algorithms, there is not a good algorithm to divide the speech model features into subspaces well. The traditional clustering algorithm uses t...

AI Technical Summary

Problems solved by technology

In the existing algorithms, there is not a good algorithm to divide the speech model features into subspaces well. The traditional clustering algorithm uses the least square method to calculate the distance between the models, which cannot effectively divide the subspaces. The existing technology has The following disadvantages:

[0003] 1. Currently based on the floating-point comparison algorithm, it takes a lot of time. After the system exceeds 2000 voice models, its real-time performance is not enough to meet business needs;

[0004] 2. Based on large-scale (more than 10,000) model feature comparisons, traditional algorithms cannot meet performance requirements;

[0005] 3. There is no effective means to divide the model features into subspaces

Images