ASR Packet Loss Correction via Signal Rectification and Score Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic Speech Recognition (ASR) systems face significant accuracy degradation and high false alarm rates due to packet loss, which occurs when data frames contain zeroes instead of actual transmitted values, especially in VoIP systems where audio is treated as network data.
Innovation Solution
The method involves rescoring phoneme probability calculations by accumulating and normalizing probabilities, deleting values for packet loss sections, replacing them with historical values, and tagging matches affected by packet loss using mel frequency cepstral coefficient features, without modifying the acoustic models generated during training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If packet loss correction is implemented at the recognition stage without modifying acoustic models, then ASR recognition accuracy improves by 2 to 4%, but the system complexity increases due to additional processing steps
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing historical values during offline training phases. These historical values are prepared in advance and stored for later use during real-time recognition, eliminating the need for complex real-time calculations when packet loss occurs. The system pre-processes data to create lookup tables and historical statistics that can be quickly applied during recognition without modifying the acoustic models.
Solution Approach 2:
The patent introduces an intermediary mechanism by using historical values and statistical models as mediators between the corrupted input signal and the recognition process. Instead of directly correcting the packet loss in the acoustic models, the system uses intermediate statistical representations (mean, variance, histograms) that bridge the gap between noisy input and reliable recognition, thereby improving accuracy without increasing model complexity.
2Measurement precision
If values for packet loss sections are replaced with historical values, then recognition accuracy improves, but computational overhead increases during real-time processing
Solution Approach 1:
The patent performs computationally intensive operations in advance during offline training. Historical values, statistical parameters (mean, variance), and probability distributions are pre-calculated and stored. During real-time recognition, the system only needs to retrieve and apply these pre-computed values, significantly reducing real-time computational overhead while maintaining high recognition accuracy.
Solution Approach 2:
The patent uses lightweight, easily computable statistical parameters (mean, variance, histograms) as disposable proxies for complex acoustic representations. These simple statistical summaries can be quickly calculated and replaced during packet loss events without requiring heavy computational resources, making the correction process computationally efficient.
3Reliability
If phoneme probability calculations are rescored with normalization, then false alarm rates are controlled, but processing time increases
Solution Approach 1:
The patent pre-computes normalization parameters and probability distributions during offline training. Statistical thresholds and normalization constants are calculated in advance and stored for rapid application during real-time processing. This allows the system to perform rescoring with normalization during packet loss events without incurring significant processing delays, as the heavy computational work has already been done beforehand.
Data Source
AI summary
A system and method are presented for the correction of packet loss in audio in automatic speech recognition (ASR) systems. Packet loss correction, as presented herein, occurs at the recognition stage without modifying any of the acoustic models generated during training. The behavior of the ASR engine in the absence of packet loss is thus not altered. To accomplish this, the actual input signal may be rectified, the recognition scores may be normalized to account for signal errors, and a best-estimate method using information from previous frames and acoustic models may be used to replace the noisy signal.