Speech Segment Grouping for TTS Concatenation Cost Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Unit selection speech synthesis systems face challenges with high computational burdens and large memory requirements due to the extensive computation needed for concatenation cost calculations, especially for complex speech patterns and high-quality synthesis, which is exacerbated by the need for pre-saving concatenation cost data.
Innovation Solution
The approach involves grouping speech segments based on their concatenation cost values and selecting representative segments for each group, allowing the concatenation cost between segments in different groups to be approximated by the representative segments, thereby reducing the amount of pre-saved concatenation cost data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If pre-saving concatenation cost data for each pair of speech segments is performed, then real-time calculation burden is reduced, but memory requirements increase to terabytes
Solution Approach 1:
The patent divides the speech segment set into multiple clusters, where each cluster contains speech segments with similar acoustic characteristics. Instead of pre-calculating concatenation costs for all possible segment pairs, the system only pre-calculates costs within each cluster and computes cross-cluster costs on-demand during runtime, dramatically reducing the pre-saved data volume while maintaining synthesis quality
Solution Approach 2:
The patent introduces cluster centers as intermediary representations. Each cluster is represented by its center speech segment, and concatenation costs between clusters are approximated by costs between their respective centers. This intermediary approach allows the system to work with compressed cluster-level data rather than full segment-level data, reducing memory requirements from terabytes to manageable sizes
2Manufacturing precision
If concatenation cost computation is performed for complex speech patterns, then synthesis quality is improved, but computational burden becomes overly burdensome
Solution Approach 1:
The patent performs preliminary clustering of speech segments based on acoustic characteristics before synthesis. This pre-organization groups similar segments together, allowing the system to quickly identify candidate segments within the same cluster and avoid exhaustive searches across all segments, thereby reducing computational complexity while maintaining the ability to handle complex speech patterns
Solution Approach 2:
The patent applies different computational strategies to different regions of the speech database. Within each cluster, full concatenation cost computation is performed to ensure high quality. Between clusters, approximate cost computation using cluster centers is sufficient. This local differentiation allows high synthesis quality where needed while reducing overall computational burden
Data Source
AI summary
Pre-saved concatenation cost data is compressed through speech segment grouping. Speech segments are assigned to a predefined number of groups based on their concatenation cost values with other speech segments. A representative segment is selected for each group. The concatenation cost between two segments in different groups may then be approximated by that between the representative segments of their respective groups, thereby reducing an amount of concatenation cost data to be pre-saved.