Speaker adaptive end-of-speech detection for conversational ai applications

Adaptive end-of-speech detection using CTC and speaker-specific thresholds enhances speech recognition accuracy and reduces latency by dynamically adjusting to individual speaking rates.

JP2026035644APending Publication Date: 2026-03-04NVIDIA CORP
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2025-11-17
Publication Date
2026-03-04

AI Technical Summary

Technical Problem

Existing systems face challenges in accurately and efficiently detecting the start and end of speech in audio data, leading to errors and high latency, especially for resource-limited devices.

Method used

Adaptive end-of-speech detection based on the speaking rate of a particular speaker, using a connectionist temporal classification (CTC) to dynamically adjust thresholds for start and end of speech detection, incorporating a sliding window and percentage of blank symbols to flag speech segments.

Benefits of technology

Improves the accuracy and reduces latency in speech detection by adapting to individual speaker characteristics, ensuring precise identification of speech segments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2026035644000001_ABST
    Figure 2026035644000001_ABST
Patent Text Reader

Abstract

Systems and methods for speaker adaptive speech end detection for conversational AI applications.SOLUTION: Determining an end-of-speech (EOS) threshold based at least in part on a speech rate, predicting, for at least one segment of a set of segments of an audio input, a set of characters representing the at least one segment, and determining an EOS corresponding to the at least one segment based at least in part on a determination that the EOS threshold is satisfied for the predicted set of characters for the at least one segment; Wherein the set of characters comprises a plurality of characters other than blanks, and wherein the EOS threshold indicates a percentage of blank characters in the set of characters.SELECTED DRAWING: FIG. 3A
Need to check novelty before this filing date? Find Prior Art