Neural Network Speech Recognition Using Latent Variable Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies face challenges in accurately transcribing speech due to distracting audio characteristics such as background noise, room reverberations, and speaker-specific features like accent and speaking style, which are not effectively filtered out.
Innovation Solution
The use of neural networks that incorporate feature vectors modeling audio characteristics and latent variables from multivariate factor analysis allows for the isolation and filtering of distracting audio features, improving speech recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional speech recognition models are used, then the system is simpler and faster, but transcription accuracy deteriorates due to inability to filter distracting audio characteristics
Solution Approach 1:
The patent segments the audio signal processing into distinct components: extracting acoustic features, extracting speaker characteristics, and combining them through neural networks. This segmentation allows each component to be optimized independently while improving overall accuracy.
Solution Approach 2:
The patent introduces additional dimensions to the traditional acoustic feature space by incorporating speaker characteristics as separate input features to the neural network. This dimensional expansion enables the system to distinguish between speech content and speaker-specific attributes, improving transcription accuracy.
2Measurement precision
If more audio features are extracted and processed, then transcription accuracy improves, but processing time and computational resources increase
Solution Approach 1:
The patent performs preliminary extraction and processing of acoustic features and speaker characteristics before they are input to the neural network. This preliminary action reduces the computational burden during the main recognition process by pre-processing the data into compact feature representations.
Solution Approach 2:
The patent extracts only the most relevant and discriminative features from the audio signal, rather than processing all available information. This selective extraction maintains transcription accuracy while significantly reducing processing time and computational requirements.
3Measurement precision
If speaker characteristics are included as input features, then speech recognition becomes more accurate, but the system becomes more complex and harder to train
Solution Approach 1:
The patent merges the extraction of acoustic features and speaker characteristics into a unified processing pipeline that feeds both types of features into a single neural network model. This merging approach simplifies the overall system architecture compared to separate processing streams while maintaining high accuracy.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for speech recognition using neural networks. A feature vector that models audio characteristics of a portion of an utterance is received. Data indicative of latent variables of multivariate factor analysis is received. The feature vector and the data indicative of the latent variables is provided as input to a neural network. A candidate transcription for the utterance is determined based on at least an output of the neural network.