Chinese Speech Recognition Pre-Training With Unlabeled Pronunciation Units
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The inefficiency and high cost of manually labeling large amounts of speech-text data for training speech recognition models, particularly for ideographic languages like Chinese, pose significant challenges in improving recognition accuracy.
Innovation Solution
A data processing system and method that pre-trains an encoder and decoder using Chinese pronunciation unit predicting tasks on unlabeled speech data, enabling the model to learn the capability of predicting Chinese pronunciation units and constructing text, thereby reducing the need for labeled data and enhancing training efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling of speech-text data is performed to improve speech recognition accuracy, then recognition accuracy is improved, but labeling cost and time consumption increase significantly
Solution Approach 1:
The system performs preliminary pre-training of the encoder and decoder models using large amounts of unlabeled speech data before actual speech recognition tasks. This preliminary action enables the model to learn general speech patterns and language structures in advance, reducing the need for extensive manual labeling later while maintaining high recognition accuracy.
Solution Approach 2:
The system uses the pre-trained models to automatically process and recognize speech data without requiring manual labeling for each new speech sample. The model serves itself by leveraging previously learned patterns to accurately transcribe new speech inputs, significantly reducing ongoing labeling requirements.
2Measurement precision
If manual labeling of speech-text data is performed to improve speech recognition accuracy, then recognition accuracy is improved, but resource consumption increases
Solution Approach 1:
The system performs preliminary pre-training of the encoder and decoder models using large amounts of unlabeled speech data before actual speech recognition tasks. This preliminary action enables the model to learn general speech patterns and language structures in advance, reducing the need for extensive manual labeling later while maintaining high recognition accuracy.
Solution Approach 2:
The system uses the pre-trained models to automatically process and recognize speech data without requiring manual labeling for each new speech sample. The model serves itself by leveraging previously learned patterns to accurately transcribe new speech inputs, significantly reducing ongoing labeling requirements.
3Productivity
If pre-training is performed on large amounts of unlabeled speech data to reduce labeling requirements, then labeling efficiency is improved, but training time and computational resources increase
Solution Approach 1:
The pre-training process is segmented into two distinct stages: first, the encoder is pre-trained separately on unlabeled speech data to learn speech representations; second, the decoder is pre-trained to perform text prediction. This segmentation allows each component to be optimized independently and enables parallel processing, reducing overall training time while improving labeling efficiency.
4Productivity
If pre-training is performed on large amounts of unlabeled speech data to reduce labeling requirements, then labeling efficiency is improved, but computational complexity increases
Solution Approach 1:
The pre-training process is segmented into two distinct stages: first, the encoder is pre-trained separately on unlabeled speech data to learn speech representations; second, the decoder is pre-trained to perform text prediction. This segmentation allows each component to be optimized independently and enables parallel processing, reducing overall training time while improving labeling efficiency.
Data Source
AI summary
A data processing system and method for a speech recognition model, a speech recognition method, a computing device and a readable storage medium. The system includes a cloud-side device and an end-side device. The cloud-side device is configured to encode, by using an encoder, sample speech data to obtain a speech feature of the sample speech data, where the encoder is pre-trained based on performing a Chinese pronunciation unit predicting task on pre-training speech data; input the speech feature into a decoder to obtain predicted Chinese text, where the decoder is pre-trained based on performing a text prediction task on a pre-training Chinese pronunciation unit; pre-train a model including the encoder and the decoder based on the predicted Chinese text and sample Chinese text, when a pre-training stop condition is met, acquire a model parameter of a speech recognition model obtained by pre-training; send the parameter to the end-side device.