Text Processing With Speech-Recognition Noise Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition models suffer from recognition errors due to environmental noise and user accents, leading to weak robustness in text processing modules, which are sensitive to noisy inputs and often fail to produce accurate results.
Innovation Solution
A method involving a noise generation model trained on correct and recognized speech data to generate noisy texts, followed by training a text processing model using these noisy texts, ensuring the generated noise reflects actual errors, thereby enhancing the model's robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If speech recognition is performed on speech data in real-world conditions, then speech recognition functionality is achieved, but recognition errors occur due to environmental noise and user accents
Solution Approach 1:
The patent applies preliminary action by pre-training the text processing model with noisy text data before actual deployment. The noise generation model creates artificially noisy training data in advance, allowing the text processing model to learn robustness to various noise patterns before encountering real-world speech recognition errors, thereby improving recognition accuracy without sacrificing adaptability
2Ease of manufacture
If word replacement based on dictionaries is used to generate noisy text, then training data with noise is obtained, but the generated noisy text cannot reflect actual speech recognition errors
Solution Approach 1:
The patent replaces the mechanical dictionary-based word replacement system with a neural network noise generation model. This model learns the actual distribution of speech recognition errors from paired clean-noisy text data and generates noisy text that accurately reflects real-world recognition mistakes, thereby improving the precision of error reflection while maintaining ease of generation through automated model inference
3Ease of operation
If the text processing module processes noisy input text, then text processing functionality is provided, but correct results often cannot be obtained due to sensitivity to noise
Solution Approach 1:
The patent converts the harmful effect of noisy input text into a benefit by using the noise generation model to create deliberately noisy training data. This allows the text processing model to learn from noisy inputs during training, transforming the previously harmful noise into a useful training signal that improves the model's ability to produce correct results even when given noisy input text
Data Source
AI summary
This application discloses example text data processing method. One example method includes obtaining a target text. The target text can then be processed based on a noise generation model to obtain a noisy text, where when the noise generation model is trained, training data of the noise generation model at least includes a first text and a second text, the first text is a correct text corresponding to speech data, and the second text is obtained by performing speech recognition on the speech data by using a first speech recognition model. A text processing model can then be trained, by using at least the noisy text as training data, to obtain a trained text processing model.