Size-Constrained Neural Network for Edge Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition (ASR) systems face challenges in deploying small-sized models on resource-constrained devices due to large model sizes and the need for training on short utterances, which is complicated by data loss during label generation for feature frames.
Innovation Solution
A method and system for automatic speech recognition in resource-constrained devices that involves receiving an audio file, transcribing it to generate an Intermediate Greedy Transcript, marking split locations, splicing the audio into segments, and training a size-constrained neural network model using these segments to recognize short bursts of words, utilizing techniques like Connectionist Temporal Classification (CTC) and Recurrent Neural Networks (RNNs).
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural network models are trained for speech recognition, then recognition accuracy is improved, but model size becomes too large to fit in cache memory of resource-constrained devices
Solution Approach 1:
The patent segments the continuous audio stream into discrete speech units (phonemes, syllables, or words) and processes them independently through the neural network. This segmentation allows the model to be trained on small, manageable units rather than requiring large contextual windows, enabling deployment on resource-constrained devices while maintaining recognition accuracy.
Solution Approach 2:
The patent changes the temporal parameter of speech processing by focusing on short-duration speech units instead of long utterances. By modifying the time dimension parameter (processing brief sounds rather than extended speech), the model achieves accurate recognition with significantly reduced complexity and size suitable for edge devices.
2Volume of stationary object
If the model is trained to recognize short bursts of words, then model size is reduced for resource-constrained devices, but data loss occurs during label generation for feature frames
Solution Approach 1:
The patent performs preliminary alignment of speech units with their corresponding labels during the data preparation phase. By pre-synchronizing the temporal boundaries of short speech bursts with their ground truth labels, the system eliminates data loss that would otherwise occur during online label generation, ensuring accurate training data without requiring manual intervention.
3Loss of time
If online speech recognition is implemented on resource-constrained devices, then low latency and privacy preservation are achieved, but limited memory and processing capacity constrain model size
Solution Approach 1:
The patent segments speech into brief units that can be processed independently and quickly, enabling low-latency recognition on resource-constrained devices. This segmentation approach allows the small model to make rapid decisions on short sound bursts without requiring extensive computation or large memory resources.
Solution Approach 2:
The patent employs a simplified model that processes only the essential features of short speech units rather than analyzing complete utterances with full contextual information. This partial processing approach reduces computational requirements and model size while maintaining sufficient accuracy for practical applications on edge devices.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
Automatic speech recognition techniques are implemented in resource constrained devices such as edge devices in internet of things where on-device speech recognition is required for low latency and privacy preservation. Existing neural network models for speech recognition have a large size and are not suitable for deployment in such devices. The present disclosure provides an architecture of a size constrained neural network and a method of training the size constrained neural network. The architecture of the size constrained neural network provides a way of increasing or decreasing number of feature blocks to achieve an accuracy-model size trade off. The method of training the size constrained neural network comprises creating a training dataset with short utterances and training the size constrained neural network with the training dataset to learn short term dependencies in the utterances. The trained size constrained neural network model is suitable for deployment in resource constrained devices.