Voice Profile Speech Endpointing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional endpointers in natural language processing systems often inaccurately segment voice inputs due to reliance on pause duration, leading to misinterpretation of user intent in multi-user conversations and background noise scenarios.
Innovation Solution
A computing device generates voice profiles based on acoustic features to identify and segment utterances from specific users, using mel-frequency cepstral coefficients, filterbank energies, or fast Fourier transform frames, and compares subsequent audio frames to determine the beginning and ending points of user-specific speech.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional endpointers use pause duration to segment voice inputs, then the segmentation process is simple, but the accuracy of identifying user intent deteriorates in multi-user conversations
Solution Approach 1:
The patent segments the voice input analysis into multiple independent components: voice profile generation from acoustic features, voice profile matching against subsequent frames, pause duration detection, and endpoint determination. This modular segmentation allows each component to specialize in one aspect of analysis, improving overall accuracy while maintaining manageable system complexity through division of labor.
Solution Approach 2:
The patent introduces voice profiles as an intermediary element between raw audio input and endpoint determination. Voice profiles serve as a mediator that captures speaker-specific acoustic characteristics, enabling the system to distinguish between different users and accurately identify whose speech should be processed, thereby resolving the contradiction between simple pause-based methods and accurate intent identification.
2Reliability
If endpointers rely on pause duration between words, then the method is easy to implement, but incorrect segmentation occurs in multi-user dialogues
Solution Approach 1:
The patent performs preliminary voice profile generation and acoustic feature extraction before conducting endpoint detection. By pre-processing the audio data to establish speaker-specific profiles and acoustic characteristics in advance, the system prepares the necessary information structures that enable reliable segmentation without requiring complex real-time analysis during the actual endpoint detection phase.
Solution Approach 2:
The patent transitions from relying solely on temporal parameters (pause duration) to incorporating spectral parameters (acoustic features, voice profiles). This parameter change from time-based to frequency-based analysis enables the system to distinguish between different speakers and background noise, significantly improving segmentation reliability while the modular implementation keeps the system manageable.
3Measurement precision
If traditional endpointing methods are used, then processing speed is fast, but misinterpretation of user intent occurs in noisy environments
Solution Approach 1:
The patent extracts and isolates speaker-specific acoustic features and voice profiles from the mixed audio signal, separating the relevant speech information from background noise and other speakers. This extraction process focuses computational resources on the most discriminative features for intent interpretation, improving accuracy while avoiding unnecessary processing of irrelevant audio components.
Solution Approach 2:
The patent applies partial action by focusing analysis on specific acoustic features and voice profile characteristics that are most relevant for speaker identification and intent interpretation. Rather than analyzing all aspects of the audio signal in detail, the system selectively processes the most informative parameters, achieving high accuracy without proportionally increasing processing time.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for speech endpointing based on a voice profile. In one aspect, a method includes the actions of receiving audio data corresponding to an utterance spoken by a particular user. The actions further include generating a voice profile for the particular user using at least a portion of the audio data. The actions further include determining in the audio data a beginning point or an ending point of the utterance based at least in part on the voice profile for the particular user. The actions further include based on the beginning point, the ending point, or both the beginning point and the ending point, outputting data indicating the utterance.