Universal Output Layer for Multilingual ASR on Limited Devices

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional multilingual automatic speech recognition (ASR) models are unsuitable for on-device applications due to their large size and computational resource requirements, as they use separate monolingual output layers for each language, leading to increased storage and computational demands.

Innovation Solution

A multilingual ASR model employing a shared monolingual output layer with language-specific wordpiece models, allowing speech recognition in multiple languages using a single output layer, thereby reducing resource consumption and enabling scalability without increasing output layer size.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate monolingual output layers are used for each language in multilingual ASR models, then language-specific speech recognition accuracy is improved, but model size and storage requirements increase significantly

Engineering Contradiction:

Improvespeech recognition accuracyVSAvoidmodel size

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent merges multiple language-specific output layers into a single shared output layer that serves all supported languages. This consolidation reduces the total number of parameters and model size while maintaining the ability to perform accurate speech recognition across multiple languages through a unified architecture.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The shared output layer is designed to be universal and multi-functional, handling speech recognition for multiple languages simultaneously. This single output layer replaces what would traditionally require separate dedicated output layers for each language, achieving resource efficiency without sacrificing language-specific performance.

Inventive Principle:

Principle #6Universality (Multi-functionality)

2Adaptability or versatility

If conventional multilingual ASR models are designed to support multiple languages, then language coverage and versatility are improved, but computational resource requirements and device compatibility deteriorate

Engineering Contradiction:

Improvelanguage coverageVSAvoidcomputational resource consumption

Core Design Contradiction:

Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

By merging language-specific processing paths into a shared output layer, the model reduces redundant computational operations. The unified architecture processes multiple languages through common computational resources, lowering the overall computational burden compared to maintaining separate processing pipelines for each language.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The shared output layer provides a universal interface that handles multiple languages with a single computational structure. This multi-functional design allows the model to support diverse languages without proportionally increasing computational resource requirements, improving efficiency for on-device deployment.

Inventive Principle:

Principle #6Universality (Multi-functionality)

3Measurement precision

If multiple separate output layers are implemented for different languages, then language-specific optimization is improved, but device storage capacity and feasibility for on-device applications worsen

Engineering Contradiction:

Improvelanguage-specific recognition precisionVSAvoidstorage capacity

Core Design Contradiction:

Measurement precisionVSVolume of stationary object

Solution Approach 1:

The patent combines multiple language-specific output layers into a single shared output layer, dramatically reducing the storage footprint. This merging eliminates redundant parameters and weights that would exist in separate output layers, making the model feasible for deployment on devices with limited storage capacity while preserving language-specific recognition capabilities.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The shared output layer serves as a universal component that handles multiple languages, replacing what would traditionally require multiple separate output layers. This universal design achieves space efficiency by storing only one set of output parameters that serve all languages, enabling on-device deployment.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS12548561B2Universal monolingual output layer for multilingual speech recognition

Publication Date: 2026.02.10 GOOGLE LLC

AI summary

A method includes receiving a sequence of acoustic frames as input to a multilingual automated speech recognition (ASR) model configured to recognize speech in a plurality of different supported languages and generating, by an audio encoder of the multilingual ASR, a higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames. The method also includes generating, by a language identification (LID) predictor of the multilingual ASR, a language prediction representation for a corresponding higher order feature representation. The method also includes generating, by a decoder of the multilingual ASR, a probability distribution over possible speech recognition results based on the corresponding higher order feature representation, a sequence of non-blank symbols, and a corresponding language prediction representation. The decoder includes monolingual output layer having a plurality of output nodes each sharing a plurality of language-specific wordpiece models.