Speech Packet Loss Concealment Using Neural Power Spectrum Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional packet loss concealment methods assume that packet losses are adjacent to previous frames, leading to poor speech quality due to perceived differences in adjacent speech signals when this assumption is not valid.
Innovation Solution
A packet loss concealment method using a neural network model to predict a lost frame power spectrum based on the speech power spectrum, allowing for restored speech data to be determined, thereby avoiding direct copying of the previous frame's pitch period waveform to the lost frame position.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the pitch period waveform signal of the previous frame is directly copied to the lost frame position, then the packet loss concealment process is simple and fast, but the speech quality deteriorates due to perceived differences between adjacent speech signals
Solution Approach 1:
The patent transforms the speech signal from time domain to frequency domain by calculating the power spectrum. Instead of directly copying the time-domain waveform, the system predicts the power spectrum parameters of the lost frame based on adjacent frames, then performs inverse transformation to generate the concealed speech signal. This parameter transformation approach enables more accurate reconstruction while maintaining computational efficiency.
Solution Approach 2:
The patent replaces the simple mechanical copying operation with a neural network-based prediction system. The neural network model analyzes the power spectrum characteristics of adjacent frames and predicts the missing frame's power spectrum, substituting the direct copying mechanism with an intelligent prediction mechanism that adapts to the actual speech signal characteristics.
2Device complexity
If conventional packet loss concealment assumes packet loss is adjacent to the previous frame, then the concealment method is straightforward, but the accuracy of speech reconstruction deteriorates when this assumption is invalid
Solution Approach 1:
The patent introduces a dynamic adaptation mechanism where the neural network model learns from the actual speech signal characteristics in the power spectrum domain. Instead of relying on static assumptions about frame adjacency, the system dynamically adjusts its prediction based on the learned patterns from training data, enabling accurate reconstruction even when packet loss is not adjacent to the previous frame.
Solution Approach 2:
The patent performs preliminary training of the neural network model using power spectrum data from speech signals. This preliminary action prepares the model to accurately predict lost frame characteristics without relying on simplistic adjacency assumptions, improving reconstruction accuracy while maintaining reasonable system complexity.
Data Source
AI summary
A packet loss concealment method includes: receiving a speech data packet; determining a speech power spectrum of speech data in the speech data packet in response to determining, according to the speech data packet, that a packet loss occurs; performing lost frame prediction on the speech power spectrum by using a neural network model, to obtain a predicted lost frame power spectrum; and determining restored speech data according to the speech power spectrum and the predicted lost frame power spectrum.