Self-Attention Speech Quality Measuring System for Real-Time Air Traffic Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech quality evaluation methods for air traffic control are time-consuming and unable to process streaming speech in real-time, and are sensitive to unvoiced parts of speech, limiting their effectiveness in real-time air traffic control scenarios.
Innovation Solution
A self-attention-based speech quality measuring method and system that processes real-time air traffic control speech data by separating voiced and unvoiced frames, using a neural network with mel spectrum auditory filtering, adaptive convolutional layers, transformer attention, and self-attention pooling to generate a predicted Mean Opinion Score (MOS) value, enabling real-time scoring and reducing the influence of unvoiced speech.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If subjective evaluation method based on MOS scoring is used, then speech quality can be evaluated intuitively, but the evaluation process takes a long time and cannot process streaming speech in real time
Solution Approach 1:
The patent replaces the manual subjective evaluation system with an automated deep learning-based objective evaluation system. The neural network model processes speech features automatically to predict MOS scores, eliminating the need for human evaluators and enabling real-time processing of streaming speech data while maintaining evaluation accuracy
Solution Approach 2:
The patent transforms the evaluation approach by changing from manual scoring parameters to automated neural network prediction parameters. The system extracts speech features (MFCC, spectral centroid, skewness, kurtosis) and feeds them into the neural network, which outputs predicted MOS scores, thereby changing the evaluation parameters from human subjective scores to machine-computed objective metrics
2Productivity
If traditional speech quality evaluation is applied, then speech data can be processed, but unvoiced parts of speech adversely affect the scoring results
Solution Approach 1:
The patent extracts and removes unvoiced speech segments from the input data before evaluation. By detecting voiced and unvoiced portions and excluding unvoiced parts from the neural network processing, the system eliminates the harmful influence of silence and non-speech segments on the MOS score prediction, ensuring that only relevant speech content is evaluated
Solution Approach 2:
The patent segments the speech signal into voiced and unvoiced portions using voice activity detection. This segmentation allows the system to process only the relevant voiced segments through the neural network while excluding unvoiced parts, thereby improving the accuracy of speech quality evaluation by focusing on actual speech content
3Measurement precision
If offline scoring system is used, then comprehensive speech quality assessment can be performed, but real-time streaming control speech cannot be processed
Solution Approach 1:
The patent implements continuous processing of streaming speech data through the neural network. The system maintains continuous evaluation capability by continuously extracting features and predicting MOS scores for incoming speech streams, enabling real-time quality assessment without the time delays associated with batch offline processing
Solution Approach 2:
The patent performs preliminary feature extraction and preprocessing of speech data before it enters the neural network for MOS prediction. By preparing the speech data in advance (extracting MFCC, spectral centroid, skewness, kurtosis features) and optimizing the neural network model, the system enables rapid real-time processing while maintaining comprehensive assessment capability
Data Source
AI summary
Disclosed are a self-attention-based speech quality measuring method and system for real-time air traffic control, including following steps: acquiring real-time air traffic control speech data and generating speech information frames; detecting the speech information frames, discarding unvoiced information frames of the speech information frames, generating a voiced long speech information frame; performing mel spectrogram conversion, attention extraction and feature fusion on the long speech information frame to obtain a predicted mos value.