Multi-Pass Speech Activity Detection for ASR Noise Robustness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition systems face performance degradations due to the inclusion of noise and non-speech acoustic events in the estimation of feature statistics and transforms, leading to mismatches in acoustic conditions, particularly in unseen noise environments.
Innovation Solution
A multi-pass speech activity detection strategy is employed, where an initial pass operates at a high miss, low-false alarm rate to discard non-speech regions, and subsequent passes focus on accurate decoding, bypassing the initial pass output to estimate feature statistics and transforms for improved acoustic modeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single-pass speech activity detection is used to identify speech regions, then the system complexity is low, but the recognition accuracy degrades when noise or non-speech events are included in feature statistics estimation
Solution Approach 1:
The speech activity detection is divided into two separate passes: a first pass that operates at high miss rate to include all potential speech regions for feature statistics estimation, and a second pass that operates at low false alarm rate to provide clean speech regions for decoding. This segmentation allows each pass to serve its specific function optimally without compromising the other.
Solution Approach 2:
The first pass performs preliminary speech activity detection to identify all potential speech regions before the actual decoding process. This preliminary action ensures that feature statistics are estimated from a comprehensive set of regions that includes all speech, preventing under-inclusion of speech segments.
Data Source
AI summary
An automatic speech recognition system and a method performed by an automatic speech recognition system are provided. The method includes performing at least two passes of speech activity detection on an acoustic utterance uttered by a speaker. The at least two passes include an initial pass and a subsequent pass. The method further includes estimating at least one of feature statistics and transforms for acoustic feature extraction and acoustic modeling based on an output of an initial pass. The method further includes performing automatic speech recognition using an output of the subsequent pass while bypassing an output of the initial pass to recognize the acoustic utterance.