Speech Recognition Pipeline for Accurate Voice Segment Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies face challenges in accurately distinguishing voice segments from noise due to computational intensity and error accumulation, leading to decreased recognition rates, particularly in methods using energy-based, statistical, and deep learning models.
Innovation Solution
A speech recognition system utilizing a processor and memory to analyze acoustic features, with a multi-stage model comprising a first model for hidden state determination, a second model for voice segment detection, and a third model for utterance content estimation, employing a combination of convolutional and transformer models for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If energy-based methods (zero-cross method) are used for VAD, then computation is relatively low, but noise segments may be erroneously recognized as voice segments
Solution Approach 1:
The patent combines multiple VAD methods (energy-based method and statistical model/deep learning method) into a unified system. The energy-based method provides initial voice segment detection with low computation, while the statistical model or deep learning model refines the detection to reduce errors from noise recognition, achieving both computational efficiency and high accuracy.
Solution Approach 2:
The patent introduces an intermediary mechanism where the energy-based method acts as a preliminary filter and the statistical model or deep learning model acts as a refinement layer. This multi-stage approach allows the system to benefit from the computational efficiency of simple methods while achieving the accuracy of complex methods, resolving the contradiction between computation and reliability.
2Reliability
If methods using statistical models or deep learning models are used for VAD, then voice segment recognition accuracy is improved, but the methods are computationally intensive
Solution Approach 1:
The patent segments the VAD process into multiple stages: first using a computationally efficient energy-based method for initial detection, then applying statistical models or deep learning models only where needed for refinement. This segmentation allows the system to achieve high accuracy without applying computationally intensive methods to all audio segments, thus reducing overall computational burden.
Solution Approach 2:
The patent applies complex statistical models or deep learning models partially - only to segments where the simple energy-based method is uncertain or where refinement is needed. This partial application of complex methods achieves the necessary accuracy improvement while avoiding the full computational cost of applying these methods throughout the entire audio signal.
3Ease of operation
If speech recognition process simply follows VAD process, then processing is straightforward, but errors due to each process accumulate and speech recognition rate decreases
Solution Approach 1:
The patent implements feedback mechanisms where the output of one process (VAD) is used to adjust and improve the next process (speech recognition). The system uses the detected voice segments from VAD to guide the speech recognition process, and errors are corrected through feedback loops that adjust parameters and reprocess segments, preventing error accumulation while maintaining process simplicity.
Solution Approach 2:
The patent performs preliminary actions in the VAD stage to prepare high-quality voice segment detection before passing to speech recognition. By accurately identifying voice segments upfront using multiple methods, the system prevents errors from propagating to the speech recognition stage, thereby maintaining both process simplicity and high recognition rates.
Data Source
AI summary
The speech recognition that is disclosed analyzes an acoustic feature for each subframe of an audio signal; provides a first model configured to determine a hidden state for each frame consisting of multiple subframes on the basis of the acoustic feature; provides a second model configured to determine a hidden state for each frame consisting of multiple subframes on the basis of the acoustic feature; and provides a third model configured to determine an utterance content on the basis of a sequence of the hidden states of each block consisting of multiple frames belonging to a voice segment.