Speech Synthesis Stress Labeling for Accurate Tone Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech synthesis systems fail to accurately incorporate stress in synthesized speech, leading to flat pronunciation and lack of expressiveness, or incorrectly apply stresses resulting in incorrect pronunciations.
Innovation Solution
A speech synthesis method and model that trains on sample texts marked with stress words and corresponding audio, enabling controlled stress pronunciations by determining stress labels at the phoneme level and incorporating target tone labels, allowing for accurate stress incorporation and tone control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional speech synthesis systems are used, then the system structure is simple, but the synthesized speech lacks stress accuracy and expressiveness
Solution Approach 1:
The patent segments the speech synthesis process into multiple components: a stress detection module that identifies stress words in text, a tone label generation module that creates tone labels based on stress patterns, and a speech synthesis module that generates audio. This segmentation allows each component to specialize in specific functions, improving stress pronunciation accuracy while maintaining manageable system complexity through modular architecture.
Solution Approach 2:
The patent applies preliminary action by pre-processing the input text to detect stress words and generate tone labels before the actual speech synthesis occurs. The stress detection module analyzes text in advance to identify which words should be stressed, and the tone label generation module pre-computes tone labels based on these stress patterns. This preliminary processing ensures that stress information is ready and accurately integrated into the final synthesized speech.
2Manufacturing precision
If stress words are manually marked in training data, then stress pronunciation accuracy improves, but data preparation time and complexity increase
Solution Approach 1:
The patent implements self-service by enabling the system to automatically detect stress words and generate tone labels without requiring manual annotation of training data. The stress detection module automatically analyzes text to identify stress patterns, and the tone label generation module automatically creates corresponding tone labels. This automation eliminates the time-consuming manual data preparation process while maintaining high stress pronunciation accuracy through the system's built-in stress detection capabilities.
Solution Approach 2:
The patent changes the parameter approach by transforming manual stress marking into automated stress detection through parameter-based methods. Instead of requiring human annotators to manually mark stress words, the system uses computational parameters and algorithms to automatically detect stress patterns in text and generate corresponding tone labels. This parameter-based automation significantly reduces data preparation time while maintaining or improving stress pronunciation accuracy through consistent, repeatable detection algorithms.
3Reliability
If the speech synthesis model incorporates stress labels and tone labels, then expressiveness and accuracy improve, but model complexity and training requirements increase
Solution Approach 1:
The patent applies the nested doll principle by integrating multiple functional layers within a unified speech synthesis model architecture. The stress detection module, tone label generation module, and speech synthesis module are nested within each other, with each module processing information and passing it to the next layer. This nested structure allows the model to incorporate stress labels and tone labels while maintaining a coherent, manageable architecture that builds complexity systematically rather than all at once.
Data Source
AI summary
The present disclosure relates to a method, apparatus, storage medium and electronic device for speech synthesis. The present disclosure enables: acquiring a text to be synthesized with a target tone label, target tone label being a tone label of a first speaker or a tone label of a second speaker; inputting the text into a speech synthesis model to obtain audio information corresponding to the text, training samples of speech synthesis model including a first sample text marked with stress words and a first sample audio of first speaker for the first sample text, and a second sample text with no stress word marked and a second sample audio of the second speaker for the second sample text, and the speech synthesis model is configured to determine stress words in the text, and generate audio information corresponding to the text according to the stress words and the target tone label.