Speech Segmentation and Model Mapping for Physiological State Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems are inadequate in accurately assessing the physiological state of a subject through speech analysis, particularly in identifying subtle changes indicative of conditions such as congestive heart failure, coronary heart disease, or psychological disorders, due to limitations in time-normalization and feature extraction methods.
Innovation Solution
A system that maps a test speech sample to a reference speech sample or a speech model, dividing both into segments based on acoustic properties, computes distances between these segments, and communicates an output indicating the physiological state based on predefined constraints and feature vectors, using dynamic time warping or deep neural networks for alignment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech samples are divided into segments based on acoustic properties and mapped using dynamic time warping, then measurement precision of physiological state assessment is improved, but device complexity increases
Solution Approach 1:
The speech sample is divided into multiple segments based on acoustic properties such as pitch, energy, and spectral characteristics. This segmentation allows for more granular analysis of different portions of speech, enabling detection of subtle physiological changes that might be missed in whole-sample analysis. The segments are created by identifying acoustic boundaries and grouping frames with similar characteristics.
Solution Approach 2:
Dynamic time warping is employed to flexibly align the test speech segments with reference speech segments, allowing for non-linear time transformations. This dynamic alignment adapts to variations in speech rate, pauses, and temporal distortions, improving the accuracy of physiological state assessment by accounting for temporal variability in speech production.
2Measurement precision
If deep neural networks are used for speech sample alignment and feature extraction, then measurement precision is improved, but use of energy increases
Solution Approach 1:
Feature extraction is performed on segmented speech samples before alignment, pre-processing the data into meaningful acoustic features such as Mel-frequency cepstral coefficients (MFCCs), pitch contours, and energy envelopes. This preliminary feature extraction reduces the dimensionality and complexity of the input to the deep neural network, lowering computational energy requirements while preserving the most relevant information for physiological state assessment.
Solution Approach 2:
The deep neural network is applied selectively to specific speech segments and feature types that are most indicative of physiological state, rather than processing all speech data uniformly. Different network architectures or processing intensities are applied to different feature types (e.g., pitch vs. spectral features) based on their relative importance for detecting specific physiological conditions.
3Reliability
If speech samples are segmented and mapped with multiple acoustic features, then reliability of physiological state assessment is improved, but difficulty of detecting and measuring increases
Solution Approach 1:
Multiple acoustic features are extracted from segmented speech samples, including pitch, energy, spectral characteristics, and temporal patterns. Each feature provides complementary information about physiological state, and their combined analysis across multiple segments increases assessment reliability through redundancy and cross-validation of findings.
Solution Approach 2:
A reference speech sample or speech model is introduced as an intermediary for comparison with the test speech sample. This reference provides a baseline for normal or known physiological states, allowing the system to detect deviations systematically. The reference acts as a mediator that translates complex multi-feature comparisons into interpretable physiological state assessments.
Data Source
AI summary
A method includes mapping, by a processor, a test speech sample, which was produced by a subject while a physiological state of the subject was unknown, to a speech model representing speech produced in a known physiological state, which speech model includes multiple acoustic states and defines allowed transitions between the acoustic states. The method further includes, based on the mapping, computing a distance between the test speech sample and the speech model, and in response to the distance, communicating an output indicating the physiological state of the subject while the test speech sample was produced. Other embodiments are also described.