Context-Aware Speech Synthesis for Consistent Multi-Sentence Audio

Solution Overview

Problem

Existing speech synthesis methods face issues with inconsistency in attributes such as style and volume among synthesized audio segments, leading to a poor user auditory experience when processing multiple sentences.

Innovation Solution

A method involving a pre-trained speech synthesis model that incorporates historical text and audio for each sentence, using features like volume, style, and emotion to ensure consistent audio synthesis across sentences.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If speech synthesis is separately performed on each sentence in the text, then processing resources of the terminal device are saved and processing speed is improved, but consistency of audio attributes such as style and volume across different sentences deteriorates

Engineering Contradiction:

Improveprocessing speedVSAvoidaudio attribute consistency

Core Design Contradiction:

ProductivityVSStability of the object's composition

Solution Approach 1:

The text is divided into multiple sentences, and speech synthesis is performed separately on each sentence. This segmentation allows for efficient processing while maintaining overall coherence through the use of historical context information from previously processed sentences.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Historical text and historical audio from previously processed sentences are stored and used as input for subsequent sentence synthesis. This preliminary action ensures that context information is available before synthesizing each new sentence, maintaining attribute consistency across the entire text.

Inventive Principle:

Principle #10Preliminary action

2Stability of the object's composition

If speech synthesis is performed on the entire text as a whole, then audio attribute consistency is improved, but processing resources and time consumption increase

Engineering Contradiction:

Improveaudio attribute consistencyVSAvoidprocessing resource consumption

Core Design Contradiction:

Stability of the object's compositionVSDevice complexity

Solution Approach 1:

The text is divided into multiple sentences, and speech synthesis is performed separately on each sentence. This segmentation allows for efficient processing while maintaining overall coherence through the use of historical context information from previously processed sentences.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The speech synthesis process maintains continuity by using historical text and historical audio from previously processed sentences as input for subsequent sentences. This ensures continuous attribute consistency throughout the entire text processing without requiring reprocessing of the entire text.

Inventive Principle:

Principle #20Continuity of useful action

3Productivity

If speech synthesis is separately performed on each sentence, then processing efficiency is improved, but auditory experience of the user deteriorates due to poor consistency

Engineering Contradiction:

Improveprocessing efficiencyVSAvoidauditory experience quality

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The text is divided into multiple sentences, and speech synthesis is performed separately on each sentence. This segmentation allows for efficient processing while maintaining overall coherence through the use of historical context information from previously processed sentences.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system uses historical audio output from previously processed sentences as feedback input for subsequent sentence synthesis. This feedback mechanism ensures that attribute consistency is maintained across sentences, improving overall auditory experience while preserving processing efficiency.

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS12456451B2Speech synthesis method, apparatus, readable medium, and electronic device

Publication Date: 2025.10.28 BEIJING YOUZHUJU NETWORK TECH CO LTD

AI summary

This disclosure relates to a speech synthesis method, an apparatus, a readable medium, and an electronic device, and the method includes: acquiring target text corresponding to each sentence in a plurality of sentences included in text to be synthesized; for the each sentence, inputting the target text corresponding to the sentence, historical text corresponding to the sentence, and historical audio corresponding to the sentence into a pre-trained speech synthesis model to acquire target audio corresponding to the sentence which is output by the speech synthesis model, wherein the historical text is target text corresponding to a previous sentence of the sentence in the text to be synthesized, and the historical audio is target audio corresponding to the historical text; and synthesizing target audio corresponding to respective sentences to obtain total audio corresponding to the text to be synthesized.