TTS Model Node Modification for Voice Signal Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current text-to-speech (TTS) models struggle with identifying and correcting errors in voice signals, leading to suboptimal speech synthesis and user dissatisfaction, as they lack effective mechanisms for node modification and error detection.
Innovation Solution
An electronic device equipped with a TTS model and an automatic speech recognition (ASR) model identifies errors in voice signals by analyzing node activities and modifying nodes based on user input, replacing or reducing weights of nodes contributing to errors, and updating the TTS model using server feedback.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a TTS model is used to generate voice signals, then speech synthesis capability is provided, but error detection and correction capability is lacking
Solution Approach 1:
The patent segments the TTS model into multiple nodes that can be independently analyzed and modified. Each node represents a specific transformation step in the text-to-speech pipeline, allowing targeted error correction without redesigning the entire model. This segmentation enables the system to identify which specific node contributes to errors and modify only that portion.
Solution Approach 2:
The patent implements a feedback mechanism where the generated voice signal is evaluated for errors, and this error information is used to iteratively modify the TTS model nodes. The system continuously monitors output quality and adjusts node weights or structures based on detected errors, creating a closed-loop system that improves reliability over time.
2Measurement precision
If node weights are reduced or nodes are replaced to correct errors, then speech synthesis accuracy is improved, but model adaptability may be compromised
Solution Approach 1:
The patent makes the TTS model dynamic by allowing node weights and structures to be adjusted based on detected errors. Rather than using fixed weights, the system dynamically modifies node parameters during operation, enabling it to adapt to different error conditions while maintaining overall model flexibility and versatility.
Solution Approach 2:
The patent applies local quality by modifying only the specific nodes that contribute to errors rather than uniformly adjusting the entire model. Each node can have different weight adjustments or replacement decisions based on its individual contribution to speech synthesis accuracy, preserving adaptability in non-error-prone regions while improving accuracy in problematic areas.
3Measurement precision
If error parts are identified through phoneme length analysis, then error detection precision is improved, but processing time increases
Solution Approach 1:
The patent applies partial action by focusing error detection efforts only on specific phoneme characteristics (length) rather than analyzing all possible voice signal properties. This selective approach provides sufficient error detection precision for TTS errors while minimizing processing time by avoiding comprehensive analysis of every signal attribute.
Data Source
AI summary
Provided is an electronic device and method of controlling same, the electronic device including: memory; and at least one processor operatively connected with the memory, wherein the at least one processor is configured to: obtain a voice signal based on a text to speech (TTS) model including a plurality of nodes, stored in the memory, wherein the voice signal corresponds to an input text, based on identifying that the voice signal includes an error, identify an error part of the voice signal which includes the identified error, identify an activity of each of the plurality of nodes related to the error part, and modify at least one node among the plurality of nodes based on the identified activity of the at least one node.