Multilingual Speech Recognition Model Expansion via Neural Machine Translation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for expanding a speech recognition model to support multiple languages require significant resources and expertise, as they necessitate redeveloping the model from a zero-base for each language, which is time-consuming and labor-intensive, and often requires native speakers or language experts.
Innovation Solution
The use of neural machine translation (NMT) to translate and adapt existing speech recognition models, allowing for the expansion of languages by translating utterances, tags, and intents, and refining the learning data to create a new model that supports multiple languages with fewer resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a speech recognition model is redeveloped from zero-base for each language, then the model can accurately recognize that specific language, but significant resources including time and manpower are consumed
Solution Approach 1:
The patent applies preliminary action by pre-translating utterances from a source language to target languages using neural machine translation before model training. This allows the speech recognition model to be adapted to multiple languages without redeveloping from scratch, significantly reducing development time while maintaining recognition accuracy through pre-prepared translated training data
Solution Approach 2:
The patent uses copying by replicating the speech recognition model structure and training process across multiple languages. Instead of creating entirely new models, the system copies the proven model architecture and populates it with translated utterances, enabling rapid multilingual deployment while preserving the accuracy characteristics of the original model
2Measurement precision
If a speech recognition model is redeveloped from zero-base for each language, then the model can accurately recognize that specific language, but substantial manpower and resources are required
Solution Approach 1:
The patent implements universality by creating a single speech recognition model framework that can handle multiple languages through neural machine translation. The system translates utterances into target languages and uses them to train the model, allowing one model to serve multiple language functions without requiring separate development resources for each language
Solution Approach 2:
The patent applies parameter changes by modifying the training data parameters (utterances) through neural machine translation rather than changing the entire model structure. This allows the model to adapt to different languages by changing the linguistic parameters of the training data while maintaining the same model architecture and reducing resource requirements
3Quantity of substance
If neural machine translation is used to expand language support, then resource utilization is reduced, but the model must translate and adapt utterances, tags, and intents effectively
Solution Approach 1:
The patent uses neural machine translation as an intermediary component that bridges source language utterances and target language training data. This mediator automatically translates and adapts the training data, reducing the need for manual language expert intervention while managing the complexity of multilingual model adaptation through an automated translation layer
Data Source
AI summary
Disclosed is a system comprising a memory storing a first natural language understanding model. The memory stores instructions that, when executed, cause a processor to: receive a request for generating a second natural language understanding model in a second language different from the first language; translate the first set of utterances into a second set of utterances in the second language; provide a second set of tags or intents to the second set of utterances; provide a user interface for receiving at least one input for modifying from among the second set of utterances or the second set of tags or intents; generate a third set of utterances and a third set of tags or intents on the basis of the input received through the user interface; and establish the second natural language understanding model including the third set of utterances and the third set of tags or intents.