Acoustic Network Model Training via BLSTM Peak Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The prediction results from bidirectional long short-term memory (BLSTM) networks in speech processing are unstable, requiring network training to achieve stable predictions.
Innovation Solution
A method and apparatus for training an acoustic network model that involves acquiring a training sample with speech files, performing peak detection using a BLSTM network, dividing speech into segments, performing semantic prediction with a CTC segmentation network, and stopping training when preset performance demands are met.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If BLSTM network is used for speech processing, then speech recognition can be performed, but prediction results are unstable
Solution Approach 1:
The patent segments the speech signal into multiple segments based on peak detection, and divides the training process into multiple stages with different loss functions. This segmentation approach allows the model to focus on different aspects of speech processing at different stages, improving prediction stability while managing training complexity through structured progression.
Solution Approach 2:
The patent performs preliminary peak detection and speech segmentation before the main training process. This preliminary action prepares the data in advance, allowing the training to proceed more efficiently with focused loss functions, thereby achieving stable predictions without excessive training complexity.
2Reliability
If network training is performed to stabilize predictions, then prediction stability improves, but training time increases
Solution Approach 1:
The patent performs preliminary peak detection and speech segmentation before training, preparing the data structure in advance. This reduces the time required during actual training iterations, allowing the model to achieve stable predictions faster without needing excessive training time.
Solution Approach 2:
The patent dynamically adjusts the loss function based on training progress, transitioning from initial loss functions to final loss functions as training advances. This dynamic adaptation allows the model to converge to stable predictions more efficiently, reducing overall training time while maintaining prediction reliability.
3Measurement precision
If speech is divided into segments for processing, then processing accuracy improves, but device complexity increases
Solution Approach 1:
The patent segments the speech signal into multiple segments based on peak detection, which improves processing accuracy by allowing focused analysis of specific speech portions. The segmentation is achieved through a dedicated module that detects peaks and divides the signal, adding complexity only where needed while maintaining overall system manageability.
Solution Approach 2:
The patent introduces an intermediary segmentation module that sits between the speech input and the main processing network. This intermediary component handles the complexity of segmentation and peak detection, simplifying the overall system architecture by isolating the complex task in a dedicated module rather than distributing it throughout the entire system.
Data Source
AI summary
Disclosed are a method and apparatus for training an acoustic network model, and an electronic device, a readable storage medium, a computer program product and a program, wherein same belong to the technical field of data processing. The method comprises: acquiring a training sample for training of an acoustic network model, wherein the training sample includes one or more speech files (S101); performing peak detection on speech included in the speech file by using a BLSTM network, so as to divide letters included in the speech file on the basis of a detected peak, thereby forming segmented speech (S102); performing semantic prediction on the segmented speech on the basis of a preset CTC segmentation network to obtain a text prediction result (S103); and after the prediction result meets a preset performance requirement, stopping the training of the acoustic network model so as to perform text prediction on input speech by using the acoustic network model (S104). The method can improve the training efficiency of an acoustic network model.