Variable Chunk Creation for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition systems face challenges in efficiently processing audio data to maximize latency and accuracy, particularly in determining optimal segmentation points for inputting data to machine learning models.
Innovation Solution
A computer-implemented method that examines audio data by time frame to identify time-dependent vocal characteristics, creating variable-length chunks based on intensity threshold crossings, and iteratively sending these chunks to a speech recognition machine learning model for processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If audio data is processed using fixed-size chunks, then processing is simpler and more consistent, but accuracy decreases because chunk boundaries do not align with natural speech breaks
Solution Approach 1:
The patent divides audio data into variable-length chunks based on vocal characteristic thresholds rather than fixed time intervals. This segmentation approach creates boundaries that align with natural speech breaks, improving recognition accuracy while maintaining manageable chunk sizes for processing.
Solution Approach 2:
The chunk creation process transitions from static fixed-size segments to dynamic variable-length segments. The system continuously monitors vocal characteristics and adjusts chunk boundaries in real-time based on detected speech patterns, allowing adaptation to varying speech rates and pause structures.
2Loss of time
If audio data is processed in smaller chunks, then latency is reduced and processing can start sooner, but more processing overhead is required to manage additional chunk boundaries
Solution Approach 1:
The system performs preliminary analysis of vocal characteristics to identify optimal chunk boundaries before full speech recognition processing begins. By detecting threshold crossings in advance, the system can prepare chunk boundaries proactively, reducing latency without requiring complex real-time boundary detection during processing.
3Manufacturing precision
If variable-length chunks based on vocal characteristics are used, then speech recognition accuracy improves by aligning with natural speech breaks, but processing complexity increases
Solution Approach 1:
The system changes the parameter used for chunk boundary determination from fixed time intervals to variable vocal characteristic thresholds. By monitoring parameters such as energy levels, pitch contours, and pause durations, the system dynamically adjusts chunk boundaries to match natural speech structure, improving accuracy without requiring overly complex processing.
Data Source
AI summary
A method, computer system, and a computer program product are provided. Audio data is received. The audio data is examined by time frame and to obtain a time-dependent vocal characteristic of the audio data. In response to the time-dependent vocal characteristic falling below an intensity threshold value at a first time point, a first chunk of the audio data is created from the audio data from a beginning time point to the first time point. The first chunk of the audio data is sent to a speech recognition machine learning model. The examining, the creating, and the sending are iteratively repeated for additional chunks of the audio data.