Multilingual Speech Recognition With Adapters for Low-Resource Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

End-to-end multilingual speech recognition systems face challenges due to data imbalance, where dominant languages with abundant training data bias the model, leading to poor performance in low-resource languages and languages with non-Latin scripts.

Innovation Solution

A transformer-based framework, called Adapt-and-Adjust (A2), incorporates language-specific and language-agnostic adapters in both the encoder and decoder, along with logit adjustments to balance class distributions, using a pre-trained mBERT model for improved language modeling and acoustic-text alignment.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a single multilingual model is trained with pooled datasets from all target languages, then parameter sharing and knowledge transfer across languages are enabled, but the model becomes biased towards dominating languages due to data imbalance

Engineering Contradiction:
Improvemultilingual capabilityVSAvoidrecognition accuracy for low-resource languages
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the training process into two distinct phases: pre-training on pooled multilingual data to establish parameter sharing, and then fine-tuning with language-specific data weighting to address individual language performance. This segmentation allows the model to benefit from multilingual knowledge transfer while preventing bias towards dominating languages during the fine-tuning stage.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent dynamically changes parameter weights during training by introducing language-specific weighting mechanisms that adjust the contribution of each language's loss function. This allows the model to adapt parameter importance based on language resource availability, ensuring low-resource languages receive appropriate attention during training.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If language-specific adapters are added to handle different languages, then performance on low-resource languages improves, but device complexity increases

Engineering Contradiction:
Improverecognition accuracy for low-resource languagesVSAvoidmodel architecture complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent implements universal adapter modules that can be applied across multiple languages with minimal modification. These adapters serve multiple functions: they adapt the base model to language-specific characteristics, enable fine-tuning on limited data, and maintain compatibility with the core multilingual architecture. This universal design reduces overall complexity compared to creating separate specialized components for each language.

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

Solution Approach 2:

The patent nests language-specific adapters within the existing multilingual transformer architecture, where adapters are inserted as additional layers within the encoder-decoder structure. This nesting allows language-specific processing to be integrated without creating separate parallel systems, thereby managing complexity through hierarchical organization.

Inventive Principle:
Principle #7Nested doll (Nesting)

Data Source

PatentEP4193356B1Systems and methods for a multilingual speech recognition framework
Publication Date: 2025.12.17 SALESFORCE INC
  • EP4193356B1 patent drawingFigure 1
  • EP4193356B1 patent drawingFigure 2A~2B
  • EP4193356B1 patent drawingFigure 3

AI summary

Embodiments described herein provide an Adapt-and-Adjust (A2) mechanism for multilingual speech recognition model that combines both adaptation and adjustment methods as an integrated end-to-end training to improve the models' generalization and mitigate the long-tailed issue. Specifically, a multilingual language model mBERT is utilized, and converted into an autoregressive transformer decoder. In addition, a cross-attention module is added to the encoder on top of the mBERT's self-attention layer in order to explore the acoustic space in addition to the text space. The joint training of the encoder and mBERT decoder can bridge the semantic gap between the speech and the text.