Text-to-Speech Language Selection for Multilingual Number Playback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In large model-based dialogue scenarios, accurately determining the language for playing text organized in sequence number in speech form is challenging due to the involvement of multiple languages in the input and output text.
Innovation Solution
A method that identifies the target language for playing sequence numbers in speech by considering both the first and second texts, using language identification based on the first text's language information and semantic information, and integrating results through a voting mechanism to enhance accuracy and reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If language identification is performed based on only the second text (output text), then the identification process is simple, but the accuracy and reliability of language identification deteriorates when multiple languages are involved
Solution Approach 1:
The patent segments the language identification task into two independent parts: identifying language from the first text (input) and identifying language from the second text (output). Each segmentation is performed separately through independent identification processes, and the results are then combined through voting. This segmentation allows each identification process to focus on specific text characteristics, improving overall accuracy while maintaining manageable complexity through modular processing.
Solution Approach 2:
The patent merges multiple language identification results (from first text and second text) through a voting mechanism to determine the final target language. By combining multiple identification outcomes rather than relying on a single source, the system achieves more reliable language identification, especially in multilingual scenarios where different text portions may indicate different languages.
2Reliability
If language identification considers both first text and second text, then the reliability of language identification improves, but the complexity of the identification process increases
Solution Approach 1:
The identification process is segmented into distinct stages: first text analysis, second text analysis, and voting aggregation. Each stage handles specific tasks independently, which improves reliability by ensuring comprehensive language analysis while controlling complexity through structured modular processing. The segmentation allows the system to systematically handle multilingual inputs without overwhelming computational complexity.
Solution Approach 2:
The voting mechanism implements a feedback loop where language identification results from both texts are aggregated and used to determine the final target language. This feedback-based aggregation ensures that the most consistent language indication across different text sources is selected, thereby improving reliability while maintaining a systematic and manageable process through iterative result refinement.
3Measurement precision
If only the second text is used for language identification, then the processing speed is fast, but the accuracy deteriorates in multilingual dialogue scenarios
Solution Approach 1:
The patent segments language identification into parallel processing of first text and second text, which can be executed simultaneously rather than sequentially. This segmentation enables the system to analyze multiple text sources for language cues without significantly increasing overall processing time, thereby maintaining productivity while improving accuracy through comprehensive multilingual analysis.
Solution Approach 2:
The system performs preliminary language identification on both texts before the final speech synthesis decision is made. By conducting language analysis in advance on all available text sources, the system prepares language identification results ahead of time, ensuring accurate target language determination without delaying the speech generation process, thus balancing accuracy and processing speed.
Data Source
AI summary
The present disclosure relates to a method of text-to-speech, a medium, and an electronic device, and the method includes: obtaining chat content, where the chat content includes a first text and a second text output by a dialogue model for the first text, and the second text includes text organized in sequence number; identifying, according to the first text and the second text, a target language for playing the sequence number in speech form; and playing the sequence number in speech form with the target language.