Parallel Speech Synthesis Using Pre-computed Hidden States
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech synthesis systems using recurrent neural networks (RNNs) face challenges in achieving real-time speech synthesis due to high computational costs and interruptions in hidden states when trying to improve sound quality and natural fluency, often resulting in low real-time rates and poor quality.
Innovation Solution
The method involves splitting text into segments and providing initial hidden states for each segment using a hidden state prediction model, allowing for parallel synthesis while maintaining continuity, thus reducing computational overhead and ensuring high-quality speech synthesis in real-time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional RNN-based speech synthesis is used to improve sound quality and natural fluency, then speech quality is improved, but computational cost increases and real-time rate decreases
Solution Approach 1:
The input text is divided into multiple segments that can be processed in parallel. Each segment is synthesized independently using separate RNN instances with initial hidden states, allowing simultaneous computation while maintaining overall speech quality. This segmentation enables parallel processing that improves real-time rate without sacrificing speech quality.
Solution Approach 2:
Initial hidden states for each segment are pre-computed based on the input text before parallel synthesis begins. This preliminary action allows each parallel RNN instance to start with the appropriate contextual information, ensuring speech quality is maintained while enabling parallel execution to improve real-time rate.
2Productivity
If parallel synthesis is attempted to improve real-time rate, then speed is improved, but hidden state continuity is interrupted and quality deteriorates
Solution Approach 1:
Initial hidden states are pre-computed for each segment based on the input text and segment position. This preliminary computation ensures that each parallel RNN instance receives the appropriate contextual information needed to maintain hidden state continuity, allowing parallel synthesis to improve real-time rate without quality loss.
Solution Approach 2:
The initial hidden states act as intermediaries that transfer contextual information between segments. By providing these pre-computed hidden states to each parallel RNN instance, the system maintains continuity of the hidden state sequence across segment boundaries, enabling quality parallel synthesis.
3Manufacturing precision
If text is processed sequentially to maintain hidden state continuity, then speech quality is maintained, but computational time increases
Solution Approach 1:
The text is segmented into independent processing units that can be synthesized in parallel. Each segment maintains its hidden state continuity through pre-computed initial hidden states, while the overall synthesis time is reduced through parallel execution of multiple segments simultaneously.
Solution Approach 2:
Initial hidden states for all segments are pre-computed before parallel synthesis begins. This preliminary action eliminates the need for sequential hidden state computation, allowing all segments to be processed in parallel while maintaining the continuity requirements for speech quality.
Data Source
AI summary
The disclosure provides a method, an apparatus, a device, and a computer-readable storage medium for speech synthesis in parallel. The method includes: splitting a piece of text into a plurality of segments; based on the piece of text, obtaining a plurality of initial hidden states of the plurality of segments for a recurrent neural network. The method further includes: synthesizing the plurality of segments in parallel based on the plurality of initial hidden states and input features of the plurality of segments.