Configurable Multilingual ASR Model for Code-Switching Accuracy

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing automatic speech recognition (ASR) systems face significant computational expense and storage waste when handling multilingual applications, particularly in scenarios requiring bilingual or trilingual support, necessitating the development of numerous specialized models.

Innovation Solution

A configurable multilingual model is developed, comprising a universal automatic speech recognition module and multiple language-specific modules, which can be dynamically configured based on user input to recognize speech in one or more target languages, reducing the need for multiple specialized models and optimizing computational resources.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If multiple specialized ASR models are built for different language combinations, then speech recognition performance for each language is improved, but computational expense and storage requirements increase significantly

Engineering Contradiction:

Improvespeech recognition performanceVSAvoidcomputational expense

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent applies universality by creating a single ASR model that can recognize multiple languages through a unified architecture. The model uses language identification to determine the spoken language and dynamically adjusts processing parameters accordingly, eliminating the need for separate specialized models for each language combination while maintaining recognition performance.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent merges multiple language-specific ASR models into a single integrated model. By combining the capabilities of multiple specialized models into one unified system that handles multiple languages simultaneously, the patent reduces the total computational resources required while maintaining the performance benefits of language-specific processing.

Inventive Principle:

Principle #5Merging (Combining)

2Reliability

If multiple specialized ASR models are built for different language combinations, then speech recognition performance for each language is improved, but storage requirements increase significantly

Engineering Contradiction:

Improvespeech recognition performanceVSAvoidstorage requirements

Core Design Contradiction:

ReliabilityVSQuantity of substance

Solution Approach 1:

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent merges multiple language-specific ASR models into a single integrated model. By combining the capabilities of multiple specialized models into one unified system that handles multiple languages simultaneously, the patent reduces the total storage requirements while maintaining the performance benefits of language-specific processing.

Inventive Principle:

Principle #5Merging (Combining)

3Adaptability or versatility

If a universal ASR model is trained to support multiple languages without language ID guidance, then model versatility is improved, but recognition accuracy for specific languages may deteriorate

Engineering Contradiction:

Improvemultilingual support capabilityVSAvoidrecognition accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent applies dynamics by making the ASR model adaptable to different languages through dynamic parameter adjustment. The model receives language identification information and dynamically modifies its processing parameters, attention mechanisms, and decoding strategies to optimize performance for the specific language being processed, thereby maintaining high accuracy across multiple languages.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent applies local quality by allowing different parts of the ASR model to be optimized for specific languages while maintaining a unified overall architecture. The model incorporates language-specific processing parameters and attention mechanisms that are activated based on the detected language, enabling each language to receive tailored processing while sharing common infrastructure.

Inventive Principle:

Principle #3Local quality

Data Source

PatentEP4364135B1Canonical training for highly configurable multilingual speech recogntion

Publication Date: 2026.02.18 MICROSOFT TECHNOLOGY LICENSING LLC

Figure 1A
Figure 1B
Figure 2A

AI summary

Embodiments are provided for building a configurable multilingual model. A computing system obtains a plurality of language-specific automatic speech recognition modules and a universal automatic speech recognition module trained on a multi-language training dataset comprising training data corresponding to each of the plurality of different languages. The computing system then compiles the universal automatic speech recognition module with the plurality of language-specific automatic speech recognition modules to generate a configurable multilingual model that is configured to selectively and dynamically utilize a sub-set of the plurality of language-specific automatic speech recognition modules with the universal automatic speech recognition module to process audio content in response to user input identifying one or more target languages associated with the audio content.