Cross-Lingual Speech Synthesis for Low-Resource ASR Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing automatic speech recognition (ASR) models struggle with low-resource languages due to limited training data and acoustic diversity, requiring separate models for each language and significant memory usage, while speech synthesis faces challenges with data sparsity and speaker variability.

Innovation Solution

A multilingual text-to-speech (TTS) model is trained with shared language and speaker embeddings, using a variational autoencoder to disentangle prosody and speaker information, and employs adversarial loss to generate diverse synthesized speech representations, leveraging high-resource languages for training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a speech recognition model is trained separately for each language, then recognition accuracy for that language is improved, but training time and computational resources increase significantly

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent combines multiple language-specific speech recognition models into a single multilingual model by generating cross-lingual speech representations. Different language inputs are transformed into a shared latent space, allowing the model to process multiple languages simultaneously without requiring separate training processes for each language, thus reducing training time while maintaining recognition accuracy.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The speech recognition model is designed to handle multiple languages through a universal cross-lingual representation framework. By using speech translation and synthesis to create language-agnostic representations, the model achieves multi-functionality, processing various languages through a single unified system rather than requiring language-specific specialized models.

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

2Measurement precision

If speech recognition models are trained on native language data only, then recognition accuracy for native speech is improved, but the model cannot recognize non-native or cross-lingual speech

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidcross-lingual recognition capability
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent introduces speech translation and synthesis as intermediary processes that convert non-native or cross-lingual speech into native language representations. This intermediary step allows the model to process diverse language inputs by first translating them into a target language for which the model has been trained, thereby extending recognition capability to cross-lingual speech while maintaining high accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The model transforms speech inputs by changing linguistic parameters through translation and synthesis operations. By converting speech from one language parameter space to another (target language), the model can recognize speech patterns it was trained on, even when the input is in a different language, thus achieving cross-lingual recognition capability.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If separate speech recognition models are used for each language, then language-specific accuracy is improved, but system complexity and deployment difficulty increase

Engineering Contradiction:
Improvelanguage-specific recognition accuracyVSAvoidsystem architecture complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges multiple language-specific recognition capabilities into a single multilingual model architecture. By using cross-lingual speech translation and synthesis to create a unified processing pipeline, the system reduces architectural complexity while preserving language-specific accuracy through shared representation learning in a common latent space.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4407605B1Using speech recognition to improve cross-language speech synthesis
Publication Date: 2026.05.13 GOOGLE LLC
  • EP4407605B1 patent drawingFigure 1
  • EP4407605B1 patent drawingFigure 2
  • EP4407605B1 patent drawingFigure 3

AI summary

Disclosed herein is a computer-implemented for training a speech recognition model, the operations comprising: obtaining a multilingual text-to-speech (TTS) model; generating, using the multilingual TTS model, a native synthesized speech representation for an input text sequence in a first language that is conditioned on speaker characteristics of a native speaker of the first language; generating, using the multilingual TTS model, a cross-lingual synthesized speech representation for the input text sequence in the first language that is conditioned on speaker characteristics of a native speaker of a different second language; generating, using the speech recognition model, a first speech recognition result for the native synthesized speech representation and a second speech recognition result for the cross-lingual synthesized speech representation; determining a consistent loss term based on the first speech recognition result and the second speech recognition result; and updating parameters of the speech recognition model based on the consistent loss term.