Scalable Dynamic Class Language Models for Uncommon Speech Terms
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems struggle to accurately transcribe unusual or uncommon terms due to limited vocabulary and training data, often misrecognizing them as more common terms.
Innovation Solution
A speech recognition system that dynamically generates class-based language models using context-specific data to identify and insert appropriate terms for uncommon words, incorporating residual unigrams and class symbols to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a generic language model with limited vocabulary is used for speech recognition, then the system complexity remains low, but the recognition accuracy for unusual or uncommon terms deteriorates
Solution Approach 1:
The language model is segmented into multiple components: a base language model for common terms, and multiple class-based language models for different categories of unusual terms (e.g., technical terms, proper nouns, domain-specific vocabulary). This segmentation allows the system to handle different term types with appropriate models, improving overall accuracy without requiring a single overly complex model.
Solution Approach 2:
The system dynamically selects and switches between different class-based language models based on the context and type of terms being recognized. Rather than using a static, monolithic model, the system adapts by activating specific class models when unusual terms are detected, making the complexity conditional and context-dependent rather than constant.
2Adaptability or versatility
If class-based language models with multiple candidate classes are used to improve recognition of unusual terms, then the adaptability to different contexts improves, but the computational complexity increases
Solution Approach 1:
Different class-based language models are designed with specialized vocabularies and probability distributions tailored to specific domains or contexts (e.g., medical terms, legal terms, technical jargon). Each class model has local expertise in its domain, allowing the system to adapt to different contexts by selecting the appropriate specialized model rather than using a single general-purpose model.
Solution Approach 2:
Multiple class-based language models are pre-trained and prepared in advance for different categories of unusual terms. When the system encounters unusual terms during speech recognition, it can immediately switch to the pre-prepared corresponding class model without needing to train or generate models on-the-fly, reducing computational complexity during runtime while maintaining high adaptability.
3Adaptability or versatility
If residual unigrams are used to represent uncommon terms in the language model, then the vocabulary coverage improves, but the precision of term recognition deteriorates due to ambiguity
Solution Approach 1:
Class-based language models serve as intermediaries between the residual unigrams (which provide broad vocabulary coverage) and the actual term recognition process. When a residual unigram is encountered, the system uses the corresponding class-based model to provide contextually appropriate candidate terms and probability distributions, resolving the ambiguity of the unigram and improving recognition precision while maintaining the vocabulary coverage benefits of using unigrams.
Data Source
Figure 1
Figure 1
Figure 2
AI summary
A method comprising receiving, at data processing hardware, a voice query spoken by a user; obtaining, by the data processing hardware, context data associated with the voice query; generating, by the data processing hardware, using a language model, a candidate transcription sequence for the voice query, the candidate transcription sequence comprising a class-based symbol that represents out-of-vocabulary terms for a language of the language model; obtaining, by the data processing hardware, a list of user-specific terms belonging to a particular class based on the context data; updating, by the data processing hardware, the language model by incorporating the obtained list of user-specific terms into the language model; and generating, by the data processing hardware, using the updated language model, a transcription for the voice query that comprises a sequence of terms including one of the user-specific terms selected from the list of user-specific terms in place of the class-based symbol.