Dynamic Offset Threshold for Speech Segment Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Voice activity detection (VAD) technologies face challenges in accurately segmenting speech in audio streams due to the difficulty in distinguishing consecutive speech segments with short frame units, leading to either excessively short or long segmentations, which negatively impact speech-to-text (STT) performance.
Innovation Solution
A method that dynamically adjusts the offset threshold based on the length of the speech segment, using two thresholds (onset and offset) to determine the start and end points of speech segments, and merges or clips segments to maintain an appropriate length for STT input, preventing excessive segment lengths and ensuring word boundaries are preserved.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a fixed threshold is used for voice activity detection, then the detection process is simple, but the speech segment length becomes inconsistent and affects STT performance
Solution Approach 1:
The patent applies dynamics by making the offset threshold variable rather than fixed. The threshold is dynamically adjusted based on the current speech segment length, allowing the system to adapt to different speech patterns while maintaining consistent segment lengths for STT processing.
Solution Approach 2:
The patent changes the parameter of the threshold from a static value to a dynamic value that depends on speech segment length. This parameter change enables the system to maintain both operational simplicity and segment length consistency by adjusting the threshold based on real-time conditions.
2Measurement precision
If a short frame unit is used for VAD, then the detection resolution is high, but it is difficult to distinguish consecutive speech segments and form coherent segments
Solution Approach 1:
The patent merges consecutive speech segments that are detected with high resolution but would otherwise be fragmented. By using dynamic thresholding based on segment length, the system combines short detected segments into coherent, longer segments suitable for STT processing.
Solution Approach 2:
The patent implements feedback by using the current speech segment length to adjust the offset threshold. This feedback mechanism allows the system to maintain high detection resolution while ensuring that consecutive segments are properly identified and merged into coherent units.
3Duration of action of moving object
If the offset threshold is increased to extend segment length, then fewer segments are produced, but speech segments may be cut off and word boundaries lost
Solution Approach 1:
The patent uses dynamic adjustment of the offset threshold based on speech segment length to prevent both excessive segmentation and word boundary loss. The threshold adapts in real-time to maintain appropriate segment lengths while preserving word boundaries.
Solution Approach 2:
The patent changes the offset threshold parameter dynamically based on the current speech segment length. This allows the system to extend segment length when needed while preventing word boundary cuts by adjusting the threshold appropriately.
4Manufacturing precision
If the offset threshold is decreased to shorten segment length, then segment precision improves, but STT input units become too short and performance deteriorates
Solution Approach 1:
The patent dynamically adjusts the offset threshold to ensure speech segments are of appropriate length for STT processing. The threshold increases as segment length increases, preventing segments from becoming too short while maintaining precision control.
Solution Approach 2:
The patent uses feedback from the current speech segment length to adjust the offset threshold. This ensures that segments are neither too short (which would harm STT performance) nor too long (which would reduce precision), maintaining optimal segment length for STT input.
Data Source
AI summary
Disclosed is a method for detecting a speech segment, which is performed by a computing device. The method may include: detecting a start point of a speech segment in an audio signal; and detecting an end point of the speech segment based on an offset threshold which is dynamically changed, and the dynamically changed offset threshold may be based on a length of the speech segment.