Entity Language Models for Speech Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face challenges in accurately transcribing topic-specific words and phrases, as they rely on general language models that are not optimized for specific domains, leading to reduced accuracy and increased manual annotation costs.
Innovation Solution
The implementation of a compositional language model that activates topic-specific entity models during speech processing, using unannotated data to improve ASR predictions by combining outputs from default and specific language models, with an attention component to weight probabilities based on context.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If general language models are used for speech recognition, then system complexity is reduced, but accuracy for topic-specific words and phrases deteriorates
Solution Approach 1:
The language model is segmented into multiple specialized entity models, each trained on specific domains or topics (e.g., sports, entertainment, news). During speech recognition, the system dynamically selects and activates only the relevant entity models needed for the current context, rather than using a single large general language model. This segmentation improves topic-specific accuracy while managing system complexity through selective activation.
Solution Approach 2:
The speech recognition system is designed to be universal by supporting multiple domains and topics through a set of interchangeable entity models. The same base speech recognition engine can handle different topics by activating appropriate entity models, making the system multi-functional without requiring separate systems for each domain.
2Measurement precision
If multiple topic-specific entity models are activated, then accuracy for domain-specific terms improves, but device complexity increases
Solution Approach 1:
The system dynamically activates entity models based on the speech context rather than loading all models simultaneously. The activation is controlled by analyzing the speech input and determining which entity models are relevant, then selectively activating only those models. This dynamic approach maintains high accuracy for domain-specific terms while reducing device complexity by keeping models inactive when not needed.
Solution Approach 2:
The system changes the activation state of entity models as a parameter based on speech context. When a topic is detected, the corresponding entity model's activation parameter is changed from inactive to active. This allows the system to adapt its complexity level dynamically based on the specific recognition task at hand.
3Measurement precision
If manual annotations are used to train language models, then model accuracy improves, but productivity decreases due to increased annotation costs
Solution Approach 1:
Entity models are pre-trained on large corpora of annotated data for specific domains before deployment. This preliminary action of training models offline on specialized data allows the system to achieve high accuracy for domain-specific terms without requiring manual annotation during runtime speech recognition, thus maintaining productivity.
Data Source
AI summary
This disclosure proposes systems and methods for leveraging entity-related language models in speech processing. A system can receive audio data corresponding to an utterance and perform automatic speech recognition (ASR) on a first portion of the audio data using a general language model. Based on the results, the system may identify a specific language model for processing a second portion of the audio data. The specific language model may include entities belonging to a common subject or class. The specific language model may, in some cases, provide better results than the general language model. While the general language model may describe a whole sentence, the specific language model may describe only a portion of a sentence. Thus, a top-level model may “activate” the specific language model when it may provide useful results. The resulting data may include results from both the general language model and the specific language model.