Speech Recognition Acoustic Model Frame Skipping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face inefficiencies due to increased calculation time as the number of frames in a speech signal increases, particularly because parallel calculations are not readily performed using neural networks for acoustic models.
Innovation Solution
The proposed method involves predicting the duration of a pronunciation in an acoustic sequence and dynamically skipping pronunciation predictions for frames corresponding to that duration, using a neural network with a label layer for pronunciation prediction and a skip layer for duration prediction, allowing for efficient processing by adjusting the size of the acoustic sequence to minimize information loss and processing time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of frames in the acoustic sequence increases to improve speech recognition accuracy, then the recognition precision improves, but the calculation time increases significantly
Solution Approach 1:
The acoustic sequence is divided into multiple frames, and the pronunciation prediction is performed selectively on specific frames rather than all frames. The sequence is segmented into prediction frames and skipped frames, where only prediction frames undergo computational processing while skipped frames are inferred from context, thereby reducing the total number of calculations required.
Solution Approach 2:
The patent implements a skipping mechanism where pronunciation predictions are omitted for certain frames that are deemed less critical or can be inferred from adjacent frames. By identifying and skipping these intermediate frames, the system reduces calculation time while preserving the essential information needed for accurate speech recognition.
2Measurement precision
If pronunciation prediction is performed for every frame to maintain accuracy, then the recognition precision is maintained, but the processing speed decreases
Solution Approach 1:
The system dynamically determines which frames require pronunciation prediction and which frames can be skipped, rather than applying a static rule to all frames. This dynamic selection allows the system to adapt to the specific characteristics of the speech signal, maintaining accuracy where needed while speeding up processing where possible.
Solution Approach 2:
Different frames are treated differently based on their local characteristics. Critical frames that contain important pronunciation information undergo full prediction processing, while less critical frames are skipped or use simplified processing, thereby optimizing the balance between accuracy and speed for each local region of the acoustic sequence.
3Measurement precision
If the acoustic sequence is processed in full to ensure complete information analysis, then the recognition accuracy is maximized, but the server computational load and costs increase
Solution Approach 1:
The patent extracts and processes only the essential frames that contain critical pronunciation information, while excluding or skipping frames that contribute less to the overall recognition accuracy. This extraction approach reduces the total computational load on the server while preserving the key information needed for accurate speech recognition.
Data Source
AI summary
A speech recognition method includes obtaining an acoustic sequence divided into a plurality of frames, and determining pronunciations in the acoustic sequence by predicting a duration of a same pronunciation in the acoustic sequence and skipping a pronunciation prediction for a frame corresponding to the duration.