Temporal-Domain Feature Extraction for ASR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems face challenges due to high computational loads and power consumption, particularly in small devices, attributed to the use of fast Fourier transform (FFT) for feature extraction, which requires significant memory and processing resources, leading to inefficient scalability and increased energy consumption.
Innovation Solution
The implementation of temporal-domain feature extraction methods, such as multi-level delta-modulation and event-driven FIR filters, eliminates the need for FFT operations, reducing computational costs and memory requirements by directly extracting events and processing only relevant data, thereby reducing the number of samples requiring mel-frequency filtering and feature extraction coefficients.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If FFT-based feature extraction is used, then speech recognition accuracy is maintained, but computational load and power consumption increase significantly
Solution Approach 1:
The patent extracts and eliminates the FFT operation from the feature extraction pipeline, retaining only the essential mel-frequency filtering functionality. By removing the computationally intensive FFT step and working directly in the temporal domain with optimized filtering, the system maintains speech recognition accuracy while dramatically reducing power consumption on resource-constrained devices.
Solution Approach 2:
The patent changes the domain of operation from frequency domain (FFT-based) to temporal domain (direct filtering), fundamentally altering how feature extraction is performed. This parameter change in the processing domain eliminates the need for complex spectral transformation while preserving the essential mel-frequency information needed for accurate speech recognition.
2Measurement precision
If FFT-based feature extraction is used, then comprehensive frequency analysis is achieved, but memory requirements and computational resources increase
Solution Approach 1:
The patent extracts only the essential frequency-selective filtering capability needed for mel-frequency analysis, eliminating the full FFT computation. By directly implementing mel-frequency bandpass filters in the temporal domain, the system achieves the necessary frequency analysis precision without the memory and computational overhead of FFT-based approaches.
Solution Approach 2:
The patent substitutes the mechanical FFT computation process with a direct temporal-domain filtering approach. Instead of transforming the signal to the frequency domain and back, the system applies optimized filters directly to the temporal signal, replacing a complex multi-step mechanical process with a simpler direct computation that achieves the same analytical goal.
3Reliability
If conventional DSP tasks are performed for feature extraction, then accurate acoustic features are obtained, but processing time increases
Solution Approach 1:
The patent extracts and removes the time-consuming FFT and windowing operations from the feature extraction pipeline, retaining only the essential filtering and energy computation steps. By eliminating redundant processing stages and working directly in the temporal domain, the system obtains accurate acoustic features with significantly reduced processing time, enabling real-time performance on mobile devices.
Data Source
AI summary
A system, article, and method provide temporal-domain feature extraction for automatic speech recognition.