Personalized Phrase Spotting in ASR via Local Cloud Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition (ASR) technologies face challenges due to computational complexities and limited computing capabilities, particularly in real-world environments with linguistic differences among individuals, making it difficult to automate complex inquiries and transactions effectively.
Innovation Solution
A method and system that detects and responds to personalized phrases by tracking their occurrences, linking them to contextual responses, and adapting language models to improve speech recognition on mobile devices, enabling localized and user-specific interactions without relying solely on cloud processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If cloud-based ASR processing is used, then speech recognition capability is improved, but computational complexity and dependency on external infrastructure increase
Solution Approach 1:
The patent segments the speech recognition system into multiple components: a cloud-based language model service that handles high-level linguistic processing, and device-local ASR components that handle acoustic processing and personalized phrase spotting. This segmentation allows the system to distribute computational complexity across different locations while maintaining high recognition accuracy.
Solution Approach 2:
The patent introduces a language model service as an intermediary between the device's ASR system and the final speech interpretation. This intermediary receives acoustic features, applies personalized language models, and returns refined hypotheses, thereby improving accuracy without requiring the entire complex processing to occur locally on the device.
2Adaptability or versatility
If personalized language models are trained, then linguistic differences among individuals are accommodated, but computational resources and training time increase
Solution Approach 1:
The patent applies local quality by implementing personalized language models that are specific to individual users' linguistic characteristics. Each user receives a customized language model trained on their personal speech patterns, vocabulary, and preferences, while the system maintains efficient resource usage by only applying these personalized models where needed in the processing pipeline.
Solution Approach 2:
The patent performs preliminary action by training personalized language models in advance and storing them for later use. During actual speech recognition operations, the pre-trained models are applied directly without requiring real-time training, significantly reducing computational energy consumption during active use while maintaining high linguistic adaptability.
3Extent of automation
If device-based ASR is implemented, then computational independence is improved, but processing capability and accuracy decrease
Solution Approach 1:
The patent segments the ASR processing pipeline into device-based components (acoustic model, feature extraction, personalized phrase spotting) and cloud-based components (language model service, vocabulary updates). This segmentation enables the device to operate autonomously with core functionality while achieving enhanced accuracy through periodic cloud interactions for model updates and complex language processing.
Solution Approach 2:
The patent implements a hybrid architecture that provides multi-functionality: the device can perform basic ASR operations independently using local models, while also supporting personalized adaptation through cloud-based language model services. This universal approach allows the system to function effectively in both offline and online modes, balancing autonomy with accuracy.
Data Source
AI summary
A method and data processing device for receiving, at a data processor, data that includes at least one personalized phrase. The method includes extracting a personalized phrase from received data. The method includes tracking, via an assigned phrase counter, each occurrence of the personalized phrase in the received data and subsequently received data. The method includes periodically comparing a value of the assigned phrase counter to pre-established count thresholds to determine when a count of the personalized phrase reaches at least one of the pre-established count thresholds. The method includes storing the personalized phrase to a phrase database and linking the personalized phrase to the one or more general phrases in the phrase database. The method includes selectively triggering a contextual response to the data and executing, by the data processor, the corresponding operation. The method includes outputting the contextual response to an output device.