Shared Monolingual Output Layer for Multilingual ASR Efficiency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional multilingual automatic speech recognition (ASR) models require separate monolingual output layers for each language, leading to increased storage and computational resources, making them unsuitable for on-device applications.

Innovation Solution

A multilingual ASR model using a shared monolingual output layer with language-specific wordpiece models, allowing speech recognition in multiple languages without significantly increasing output layer size, thereby optimizing resource efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate output layers are used for each language, then language-specific recognition accuracy is improved, but device complexity and training efficiency deteriorate

Engineering Contradiction:
Improvelanguage-specific recognition accuracyVSAvoidoutput layer complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies universality by designing a single output layer that serves multiple language recognition functions. Instead of creating separate output layers for each language, the system uses one universal output layer with language-specific embedding vectors to achieve multilingual recognition, thereby reducing device complexity while maintaining language-specific accuracy

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent merges multiple language-specific processing paths into a unified architecture. By combining language identification, language embedding generation, and speech recognition into a single integrated system with one output layer, it eliminates redundancy and simplifies the overall device structure while preserving multilingual capabilities

Inventive Principle:
Principle #5Merging (Combining)

2Measurement precision

If multiple language-specific models are trained separately, then language-specific performance is improved, but training time and computational resources deteriorate

Engineering Contradiction:
Improvelanguage-specific performanceVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent combines multiple language-specific training processes into a single unified training operation. By training one model with a universal output layer on multilingual data simultaneously, it achieves language-specific performance without the need for separate training runs, significantly reducing training time and computational resource consumption

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The universal output layer enables a single model to perform multiple language recognition tasks simultaneously. This multi-functional approach allows the system to achieve language-specific performance through unified training, eliminating the time loss associated with training multiple separate models

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If language identification is performed before recognition, then recognition accuracy is improved, but processing speed deteriorates

Engineering Contradiction:
Improverecognition accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The patent merges language identification and speech recognition into a single simultaneous processing operation. The system performs both language identification and recognition in parallel using the same neural network passes, eliminating the sequential processing bottleneck while maintaining recognition accuracy through language-specific embedding vectors

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The system maintains continuous processing by performing language identification and speech recognition simultaneously without interruption. The unified architecture allows both tasks to proceed in parallel throughout the processing pipeline, preventing the speed reduction that would occur with sequential processing steps

Inventive Principle:
Principle #20Continuity of useful action

4Measurement precision

If separate processing pipelines are used for each language, then language-specific accuracy is improved, but system complexity deteriorates

Engineering Contradiction:
Improvelanguage-specific accuracyVSAvoidprocessing pipeline complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies universality by creating a single processing pipeline that handles multiple languages through language-specific embedding vectors. This universal pipeline eliminates the need for separate processing chains for each language, reducing system complexity while maintaining language-specific accuracy through the embedded language identifiers

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The system merges multiple language-specific processing pipelines into one unified pipeline. By combining language embedding generation, language identification, and speech recognition into a single integrated processing flow, it achieves language-specific accuracy without the complexity of maintaining separate processing chains for each language

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4578006B1Universal monolingual output layer for multilingual speech recognition
Publication Date: 2026.05.06 GOOGLE LLC
  • EP4578006B1 patent drawingFigure 1
  • EP4578006B1 patent drawingFigure 2A
  • EP4578006B1 patent drawingFigure 2B

AI summary

A method (500) includes receiving a sequence of acoustic frames (100) as input to a multilingual automated speech recognition (ASR) model (200) configured to recognize speech in a plurality of different supported languages and generating, by an audio encoder (204) of the multilingual ASR, a higher order feature representation (212, 222) for a corresponding acoustic frame. The method also includes generating, by a language identification (LID) predictor (230) of the multilingual ASR, a language prediction representation (232) for a corresponding higher order feature representation. The method also includes generating, by a decoder (240) of the multilingual ASR, a probability distribution (252) over possible speech recognition results based on the corresponding higher order feature representation, a sequence of non-blank symbols (121), and a corresponding language prediction representation. The decoder includes monolingual output layer (400) having a plurality of output nodes (410) each sharing a plurality of language-specific wordpiece models (420).