Speech Synthesis Using Spliced Text and Spectral Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech synthesis technologies produce unnatural and low-quality sounds with slow synthesis speeds, especially when dealing with a small amount of training data.
Innovation Solution
The integration of an auto-regressive acoustic model with a neural network vocoder, where text features are spliced with spectral features from previous time steps to enhance naturalness and quality, and the use of convolutional neural networks for smoothness processing, reduces calculation and improves synthesis speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If existing speech synthesis methods are used, then speech can be synthesized, but the synthesized sound is unnatural, low quality, and slow
Solution Approach 1:
The speech synthesis system is segmented into distinct functional modules: text preprocessing module, acoustic model module, and vocoder module. Each module handles specific tasks independently, allowing for optimized processing at each stage while maintaining overall system efficiency and quality
Solution Approach 2:
Text preprocessing is performed in advance to extract and prepare text features before they are fed into the acoustic model. This preliminary action includes text normalization, segmentation, and feature extraction, which prepares the data for more efficient and accurate speech synthesis processing
2Manufacturing precision
If existing speech synthesis methods are used, then speech can be synthesized, but the synthesized sound is unnatural and low quality
Solution Approach 1:
The system transforms text information into acoustic parameters through the acoustic model, which generates phoneme-level acoustic features. By changing the representation parameters from raw text to acoustic parameters, the system achieves more natural speech synthesis while managing model complexity through efficient parameter transformation
3Productivity
If existing speech synthesis methods are used, then speech can be synthesized, but the synthesis speed is slow
Solution Approach 1:
The system replaces traditional mechanical speech synthesis methods with neural network-based acoustic modeling. The acoustic model uses deep learning to directly generate acoustic parameters from text, eliminating the need for complex rule-based systems and achieving both high speed and high quality speech synthesis
Data Source
AI summary
A speech synthesis method, a speech synthesis device, and an electronic apparatus are provided, which relate to a field of speech synthesis. Specific implementation solution is the following: inputting text information into an encoder of an acoustic model, to output a text feature of a current time step; splicing the text feature of the current time step with a spectral feature of a previous time step to obtain a spliced feature of the current time step, and inputting the spliced feature of the current time step into an decoder of the acoustic model to obtain a spectral feature of the current time step; and inputting the spectral feature of the current time step into a neural network vocoder, to output speech.