Neural Network Speech Recognition via Keyword Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition technologies face high error rates and resource-intensive computing requirements, especially when dealing with variable speech instructions and noise interference, making them unsuitable for embedded systems and user-friendly applications.
Innovation Solution
A semantic recognition method and device that generate a spectrogram of a speech signal, detect keywords using a neural network model, and determine semantic intentions based on keyword categories, thereby increasing flexibility and reducing computing resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If NLU technique is applied to perform word segmentation and word rearrangement for flexible speech recognition, then adaptability of speech instruction is improved, but computing resource consumption increases significantly
Solution Approach 1:
The patent segments the speech recognition process into two stages: first using a lightweight acoustic model to extract potential keyword candidates from the speech signal, then using a language model to filter and identify the actual keywords. This segmentation allows the system to achieve flexible speech recognition without requiring the full computational power of traditional NLU techniques, as each stage processes only relevant information.
Solution Approach 2:
The patent extracts only the essential keyword detection function from the complex NLU technique. Instead of applying full natural language understanding processing to every speech input, the system extracts and processes only the keyword candidates that are most relevant to the speech instruction, significantly reducing computing resource requirements while maintaining adaptability.
2Measurement precision
If NLU technique with huge acoustic and language databases is used, then semantic recognition accuracy is improved, but device complexity increases making it unsuitable for embedded systems
Solution Approach 1:
The patent divides the large-scale language understanding task into two smaller, more manageable components: an acoustic model for keyword candidate extraction and a language model for keyword identification. This segmentation reduces the complexity of each individual component, making them suitable for embedded systems while collectively achieving high semantic recognition accuracy.
Solution Approach 2:
The patent introduces an intermediary keyword candidate extraction stage that bridges the gap between simple acoustic recognition and complex language understanding. This intermediary layer processes speech signals to generate a reduced set of potential keywords, which then feed into the language model, effectively mediating between the speech input and the final semantic interpretation.
3Use of energy by moving object
If fixed speech length and fixed word order are required for speech instruction, then computing resource consumption is reduced, but error rate increases when users speak naturally
Solution Approach 1:
The patent implements a dynamic keyword detection approach where the system adapts to different speech patterns and word orders. The acoustic model dynamically identifies potential keywords regardless of their position in the speech instruction, and the language model dynamically determines which candidates are actual keywords based on contextual information, allowing the system to handle variable speech structures efficiently.
Data Source
AI summary
A semantic recognition method and a semantic recognition device are provided. A spectrogram of a speech signal is generated. At least one keyword of the spectrogram is detected by inputting the spectrogram into a neural network model. A semantic category to which each of the at least one keyword belongs is distinguished. A semantic intention of the speech signal is determined according to the at least one keyword and the semantic category of the at least one keyword.