Single-Module ASR Denormalization with Contextual Rule Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition (ASR) systems require multiple modules to handle different denormalization tasks, leading to increased cost and latency, and lack customization for individual user preferences and contexts.
Innovation Solution
A single ASR module dynamically applies a contextualized list of denormalizers based on context metadata, such as user preferences and application type, to personalize speech recognition results without needing multiple modules.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple ASR modules are used to handle different denormalization tasks, then the system can provide specialized processing for different applications, but the cost and latency increase
Solution Approach 1:
The patent implements a single ASR module that performs multiple denormalization tasks by dynamically selecting and applying different denormalizer rules based on context metadata. This universal module can handle various application-specific requirements (voice search, dictation, captions, etc.) without requiring separate specialized modules, thereby reducing system complexity while maintaining adaptability.
Solution Approach 2:
The system dynamically adjusts the denormalization process by selecting different denormalizer rules based on real-time context metadata (application type, user preferences, speech characteristics). This dynamic adaptation allows a single module to provide specialized processing capabilities for different applications without requiring multiple static modules, resolving the contradiction between versatility and complexity.
2Adaptability or versatility
If multiple ASR modules are maintained for different applications, then application-specific optimization is achieved, but latency increases due to module switching and access overhead
Solution Approach 1:
The patent segments the denormalization process into independent, modular denormalizer rules that can be selectively applied within a single ASR module. Each denormalizer handles specific tasks (capitalization, punctuation, number formatting) and can be independently selected based on context, eliminating the need to switch between complete ASR modules and reducing latency.
Solution Approach 2:
The system pre-loads multiple denormalizer rules into memory within the single ASR module, allowing rapid selection and application based on context metadata without requiring module instantiation or loading. This preliminary preparation eliminates switching overhead and reduces latency while maintaining application-specific optimization.
3Device complexity
If a single ASR module is used for all applications, then cost and latency are reduced, but customization for individual user preferences and contexts is limited
Solution Approach 1:
The patent applies local quality by allowing different denormalizer rules to be selectively applied to different portions of the speech recognition output based on context metadata. Each rule can be independently enabled or disabled for specific applications or user preferences, providing customization without requiring multiple complete modules. This maintains a single module structure while achieving tailored output for different users and contexts.
Solution Approach 2:
The system changes parameters (denormalizer selection, rule priorities, processing depth) based on context metadata such as application type, user preferences, and speech characteristics. This parameter adjustment allows a single ASR module to adapt its behavior to different scenarios, providing customization capability without increasing module complexity.
4Device complexity
If denormalization rules are applied in a fixed sequence, then the processing pipeline is simple, but the output cannot be optimized for specific user states and preferences
Solution Approach 1:
The patent implements dynamic sequencing of denormalizer rules based on context metadata. The system determines the optimal execution order and selection of denormalizers according to the specific application, user preferences, and speech characteristics, replacing fixed sequencing with adaptive sequencing. This maintains relative simplicity while achieving user-specific optimization through dynamic rule selection and ordering.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A method (600) includes receiving a speech input (104) from a user and obtaining context metadata (110) associated with the speech input. The method also includes generating a raw speech recognition result (312) corresponding to the speech input and selecting a list of one or more denormalizers (352) to apply to the generated raw speech recognition result based on the context metadata associated with the speech input. The generated raw speech recognition result includes normalized text. The method also includes denormalizing the generated raw speech recognition result into denormalized text (322) by applying the list of the one or more denormalizers in sequence to the generated raw speech recognition result.