Pause-Based Conversational Turn Segmentation for Transcript Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech-to-text applications struggle to accurately segment conversational turns, particularly when speakers speak in parallel, leading to inaccurate analysis of transcripts and disrupting the natural flow of conversations.
Innovation Solution
A system that segments transcripts into primary and secondary turns by isolating listener utterances during a speaker's turn, using a natural turn model to preserve the naturalistic flow of conversations by separating listener utterances as secondary turns and maintaining speaker turns as primary, with adjustable pause thresholds and speech type identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the speech to text application inserts a turn whenever someone speaks, then it captures all speaker contributions, but it interrupts the speech of the primary speaker and creates inaccurate speaking turns
Solution Approach 1:
The patent segments conversational turns into primary turns and secondary turns based on pause duration. Primary turns are identified when a speaker pauses for longer than a threshold duration, while secondary turns capture shorter interruptions. This segmentation resolves the contradiction by creating a hierarchical structure that maintains accuracy without requiring complex real-time decision logic.
Solution Approach 2:
The system performs preliminary classification of turns by analyzing pause durations before finalizing the transcript structure. By pre-identifying which turns are primary versus secondary based on temporal characteristics, the system avoids complex real-time negotiations about turn boundaries and achieves accurate segmentation through predetermined criteria.
2Ease of operation
If the speech to text application segments the transcript into speaking turns for respective speakers, then it improves readability, but it disrupts the natural flow of conversations when speakers speak in parallel
Solution Approach 1:
The patent applies different quality characteristics to different parts of the transcript. Primary turns are formatted to emphasize the main speaker's contribution and natural pauses, while secondary turns are formatted differently to indicate shorter interruptions. This local differentiation maintains readability while preserving the natural conversational flow by visually distinguishing between different types of speech acts.
Solution Approach 2:
The system adds a temporal dimension to the transcript segmentation by using pause duration as the key criterion for distinguishing primary and secondary turns. This dimensional approach allows the transcript to represent both readability (through structured turns) and natural flow (through temporal-based classification) simultaneously, resolving the contradiction between these two requirements.
3Quantity of substance
If the system creates a turn for every speaker contribution including brief interjections, then it captures all speech events, but it fragments the primary speaker's turn and reduces analysis accuracy
Solution Approach 1:
The patent implements a dynamic classification system where the status of a speech event (primary turn vs. secondary turn) is determined by its temporal characteristics relative to other events. Rather than using a fixed rule that every speech event creates a separate turn, the system dynamically adjusts turn boundaries based on pause duration thresholds, allowing brief interjections to be absorbed into primary turns when appropriate while still capturing all speech events.
Solution Approach 2:
The system changes the parameter used for turn segmentation from a simple presence-based criterion (any speech creates a turn) to a duration-based criterion (pause length determines turn boundaries). This parameter change allows the system to capture all speech events while maintaining analysis accuracy by using the temporal parameter to distinguish between meaningful turn boundaries and minor interruptions.
Data Source
AI summary
In some embodiments, a method receives a first transcript that includes a first speaker and a second speaker. A boundary of a primary turn between the first speaker and the second speaker is determined in the first transcript. The method compares a time in which the first speaker paused to a threshold. When the threshold is met, speech by the second speaker is determined that should be labeled with a first label as the primary turn. When the threshold is not met, speech by the second speaker is determined that should be labeled with a second label as a secondary turn. The method transforms the first transcript into a second transcript based on whether speech is labeled with the first label or the second label. The second transcript is analyzed to generate an analysis of primary turns between the first speaker and the second speaker.