Speaker Turn Detection at Word Boundaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current diarization systems face issues with false positives, noise from non-verbal acoustic cues, and sub-optimal statistics estimation due to the lack of constraints on speaker turn detection, leading to poor automated speech recognition performance.
Innovation Solution
The proposed approach uses Voice Activation Detection (VAD) followed by Automatic Speech Recognition (ASR) to filter out non-speech segments and detect speaker turns at word boundaries, reducing false positives and improving statistics estimation by constraining speaker turns to whole words, thereby enhancing diarization and ASR performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a sliding window of fixed length is used to detect speaker turns, then the system can identify potential speaker changes, but it creates a big number of false positives and noisy statistics estimation
Solution Approach 1:
The patent changes the parameter of speaker turn detection from continuous time points to discrete word boundary points. By constraining potential speaker turns to occur only at word boundaries rather than at any time within a sliding window, the system reduces false positives while maintaining detection accuracy. This parameter change transforms the detection space from continuous to discrete, eliminating spurious detections that occur within words.
2Adaptability or versatility
If speaker turns are detected without constraints, then the system can find all potential turns, but the ASR performance deteriorates due to discontinuities in speech flow
Solution Approach 1:
The patent applies a constraint parameter that limits speaker turn detection to word boundaries only. This parameter change ensures that speaker turns do not occur in the middle of words, thereby maintaining speech flow continuity and preventing ASR performance deterioration while still capturing all meaningful speaker transitions that occur at word-level boundaries.
3Adaptability or versatility
If the length of sub-segments is not constrained, then the system can flexibly segment speech, but sub-optimal statistics estimation affects subsequent processes
Solution Approach 1:
The patent changes the parameter of segment length from unconstrained to constrained by word boundaries. By requiring that speaker turns occur at word boundaries, the system ensures that each sub-segment contains complete words, which provides sufficient and optimal data for statistics estimation while maintaining flexibility in segmenting speech into meaningful linguistic units.
Data Source
AI summary
An approach is provided that receives an audio stream and utilizes a voice activation detection (VAD) process to create a digital audio stream of voices from at least two different speakers. An automatic speech recognition (ASR) process is applied to the digital stream with the ASR process resulting in the spoken words to which a speaker turn detection (STD) process is applied to identify a number of speaker segments with each speaker segment ending at a word boundary. A speaker clustering algorithm is then applied to the speaker segments to associate one of the speakers with each of the speaker segments.