Offline Speech Recognition Syllable Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Offline speech recognition systems face challenges in accuracy due to limited computing and storage resources, leading to lower recognition results, while online systems compromise on privacy and reliability due to network dependency.
Innovation Solution
The method employs a syllable-based approach, using a syllable-based end-to-end SMLTA model and a syllable-based LSTM language model, with 4-bit quantization to reduce model size, and a streaming transformer model to transform syllable recognition results into readable text, effectively improving accuracy without increasing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If offline speech recognition is implemented on resource-constrained devices, then privacy and reliability are improved, but recognition accuracy deteriorates due to limited computing and storage resources
Solution Approach 1:
The speech recognition task is segmented into two independent stages: syllable recognition stage and text transformation stage. The syllable recognition model processes acoustic features to identify syllables, while the text transformation model converts syllable sequences into readable text. This segmentation allows each model to be optimized independently for resource-constrained devices while maintaining overall recognition accuracy.
Solution Approach 2:
The patent applies 4-bit quantization to reduce the precision of model parameters from standard 32-bit floating point to 4-bit integers. This parameter change significantly reduces the storage requirements and computational complexity of the models, enabling them to run on resource-constrained devices while maintaining acceptable recognition accuracy through careful quantization-aware training.
2Quantity of substance
If model size is reduced to fit resource-constrained devices, then device compatibility is improved, but recognition accuracy deteriorates
Solution Approach 1:
By dividing the speech recognition system into separate syllable recognition and text transformation models, each model can be kept compact in size while collectively achieving high accuracy. The syllable recognition model focuses on acoustic pattern matching, and the text transformation model handles linguistic conversion, allowing both to be optimized for small device memory constraints.
Solution Approach 2:
The application of 4-bit quantization directly addresses the model size issue by reducing the memory footprint of model parameters by approximately 8-fold compared to standard 32-bit floating point representations. This parameter precision reduction enables deployment on resource-constrained devices while quantization-aware training techniques maintain recognition accuracy.
3Measurement precision
If online speech recognition is used to achieve high accuracy, then recognition accuracy is improved, but privacy and reliability deteriorate due to network dependency
Solution Approach 1:
The patent enables the device to perform speech recognition independently without requiring external server assistance. By implementing both syllable recognition and text transformation capabilities locally on the device, the system achieves self-sufficient offline operation, eliminating network dependency while maintaining privacy and reliability.
Data Source
AI summary
The present disclosure provides a method of recognizing speech offline, electronic device, and a storage medium, relating to a field of artificial intelligence such as speech recognition, natural language processing, and deep learning. The method may include: decoding speech data to be recognized into a syllable recognition result; transforming the syllable recognition result into a corresponding text as a speech recognition result of the speech data.