Finite State Transducer for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems face increased processing load and memory requirements due to the need to determine utterance rates and use multiple models for recognizing speeches with varying rates, such as faster-than-normal and slower-than-normal utterances.
Innovation Solution
A generation device generates a weighted finite state transducer (WFST) that can handle slow utterances by embedding loops in the WFST, allowing for recognition of phonetic symbols that can be uttered slowly, enabling speech recognition with a single model that does not require utterance rate determination, thereby reducing processing load and memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple utterance-rate-dependent models are used to recognize speeches with varying rates, then recognition accuracy for both fast and slow utterances is improved, but processing load and memory requirements increase
Solution Approach 1:
The patent merges multiple utterance-rate-dependent models into a single model by integrating fast-utterance handling capabilities into the normal speech recognition model. This is achieved by adding specific transitions and states to the finite state transducer (FST) that enable it to handle both normal and fast utterances without requiring separate models, thereby reducing processing load while maintaining recognition accuracy.
Solution Approach 2:
The patent makes the single speech recognition model universal by enabling it to handle multiple utterance rates (normal and fast) through the FST structure. The FST is designed with transitions that can accommodate variable utterance rates, allowing the same model to function effectively for different speech speeds without requiring multiple specialized models.
2Reliability
If multiple utterance-rate-dependent models are used to recognize speeches with varying rates, then recognition accuracy for both fast and slow utterances is improved, but memory area requirements increase
Solution Approach 1:
The patent combines multiple models into a single unified model structure. Instead of storing separate models for normal, fast, and slow utterances, the system uses one FST-based model that incorporates handling for all utterance rates through its transition structure, significantly reducing the memory area required to store multiple models.
Solution Approach 2:
The single speech recognition model is designed to be universal across different utterance rates. The FST structure includes transitions that can adapt to variable utterance speeds, allowing the same model to be used for recognizing both fast and slow utterances, thereby eliminating the need for multiple model copies in memory.
3Reliability
If utterance rate determination operations are performed to handle varying speech rates, then recognition accuracy is improved, but processing load increases
Solution Approach 1:
The patent performs preliminary action by pre-configuring the FST with transitions that can handle variable utterance rates without requiring real-time determination. The FST structure is designed in advance to accommodate fast utterances through alternative transition paths, eliminating the need for runtime utterance rate classification and reducing processing load during actual speech recognition.
Solution Approach 2:
The patent extracts the utterance rate determination operation from the recognition process. Instead of determining utterance rate as a separate preprocessing step, the FST structure directly handles variable rates through its transition design, removing the need for explicit utterance rate classification and reducing overall processing load.
Data Source
AI summary
A generation device includes a receiving unit and a generating unit. The receiving unit receives a model representing correspondence between one or more phonetic symbols and one or more words. The generating unit generates a first finite state transducer based on the model, the first finite state transducer at least including, as outgoing transitions from a first state representing transition destination of a first transition which has a first phonetic symbol of a predetermined type as input symbol, a second transition that has a second phonetic symbol, which is different than a particular symbol representing part or whole of input symbol of the first transition, as input symbol, and a third transition that has a third phonetic symbol, which represents the particular symbol or silence, as input symbol.