Speech Recognition Segmentation Using Linguistic Boundary Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition systems face issues with accurate audio segmentation, leading to incomplete sentence transcriptions due to time-out-based logic, which degrades the quality of output and causes confusion in downstream processing.
Innovation Solution
The system continuously decodes streaming audio data, determines linguistic boundaries, and applies punctuation only when a complete sentence is identified, then waits for a pre-determined number of newly decoded words to validate the punctuation before outputting the segmented audio, ensuring accurate and complete sentence transmission.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If time-out based logic is used for audio segmentation, then the system can process audio continuously without complex processing, but the segmentation accuracy deteriorates causing incomplete sentence transcriptions
Solution Approach 1:
The patent segments audio into utterances based on linguistic boundaries detected by the ASR system rather than using fixed time-out intervals. The segmentation is driven by detecting when a complete sentence has been transcribed, allowing flexible segmentation that adapts to natural speech patterns while maintaining continuous processing capability.
Solution Approach 2:
The system uses feedback from the ASR transcription output to control segmentation. The punctuation engine monitors the transcribed text and uses linguistic analysis of the transcription to determine when segmentation should occur, creating a feedback loop that improves segmentation accuracy based on actual transcribed content rather than relying solely on timing.
2Measurement precision
If neural network based models are used to combine acoustic and linguistic signals for segmentation, then segmentation quality improves, but the system complexity increases and over-segmentation occurs
Solution Approach 1:
The patent applies segmentation at the linguistic level by identifying sentence boundaries in the transcribed text rather than segmenting audio based on acoustic patterns alone. This linguistic segmentation approach simplifies the system by leveraging existing ASR output while achieving accurate segmentation without complex neural network models.
Solution Approach 2:
The punctuation engine acts as an intermediary between the ASR system and the final output. It receives raw transcribed text, applies linguistic rules and heuristics to identify sentence boundaries, and produces properly segmented and punctuated output. This intermediary layer simplifies the overall system architecture by separating the ASR function from the segmentation function.
3Measurement precision
If audio is segmented at linguistic boundaries, then complete sentences are produced, but the processing time increases due to validation requirements
Solution Approach 1:
The system performs preliminary transcription of audio segments and then applies linguistic analysis to identify boundaries. By pre-transcribing the audio and then analyzing the transcription for linguistic patterns, the system can quickly identify sentence boundaries without requiring complex real-time analysis during audio processing.
Solution Approach 2:
The punctuation engine dynamically adjusts segmentation parameters based on the transcribed text content. It monitors linguistic features such as punctuation marks, capitalization patterns, and sentence structure to determine optimal segmentation points, allowing flexible processing time adjustment based on the specific characteristics of each audio segment.
Data Source
AI summary
Systems are configured to obtain streaming audio data comprising language utterances, continuously decode the streaming audio data in order to generate decoded streaming audio data and determine whether a linguistic boundary exists within an initial segment of decoded streaming audio data. When a linguistic boundary is determined to exist, the systems apply a punctuation at the linguistic boundary and output a first portion of the initial segment of the streaming audio data ending at the linguistic boundary while refraining from outputting a second portion of the initial segment which is located temporally subsequent to the first portion of the initial segment. Systems are also configured to delay the output until predetermined punctuation validation processes have been performed.