Language-Independent Neural Network for Multi-Lingual Speech Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional automatic speech recognition (ASR) systems require language-dependent resources, which are error-prone and increase development effort, complicating the development of multi-lingual recognition systems.

Innovation Solution

A language-independent neural network is constructed using multiple speech corpora to recognize spoken utterances and identify languages jointly, eliminating the need for language-dependent resources by using a hybrid attention/CTC architecture for end-to-end speech recognition.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If language-dependent resources (pronunciation dictionaries, word segmentation) are used in conventional ASR systems, then speech recognition accuracy can be improved, but development effort and system complexity increase significantly

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts and removes language-dependent resources (pronunciation dictionaries, word segmentation models, phoneme representations) from the ASR system. By eliminating these intermediate representations and directly mapping acoustic features to character sequences, the system achieves language-independent operation while maintaining recognition accuracy through the attention mechanism and CTC loss function.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent creates a universal ASR system that can process multiple languages simultaneously using a single neural network architecture. The model uses a shared character set across languages and language identification tokens to handle different languages without requiring separate language-specific modules, achieving multi-functionality in a single system.

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

2Measurement precision

If language-dependent resources are hand-crafted for each language, then recognition performance can be optimized, but development time and effort greatly increase

Engineering Contradiction:
Improverecognition performanceVSAvoiddevelopment time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent enables the system to automatically adapt to new languages without manual resource creation. The neural network learns language-specific patterns automatically from raw speech data and text transcripts through self-supervised learning, eliminating the need for hand-crafted pronunciation dictionaries and word segmentation rules for each language.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent performs preliminary training on multi-lingual data to pre-learning language-invariant acoustic features and character-level representations. This preliminary action allows the model to quickly adapt to new languages with minimal additional training, reducing development time while maintaining performance.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If separate language-dependent modules are used for each language, then language-specific accuracy can be improved, but device complexity and resource requirements increase

Engineering Contradiction:
Improvelanguage-specific accuracyVSAvoidresource requirements
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent merges multiple language-specific processing paths into a single unified neural network. By combining language identification, acoustic feature processing, and character sequence generation into one end-to-end model, the system reduces resource requirements while maintaining language-specific accuracy through the attention mechanism that dynamically focuses on relevant features for each language.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3516650B1Method and system for training a multi-language speech recognition network
Publication Date: 2020.05.13 MITSUBISHI ELECTRIC CORP
  • EP3516650B1 patent drawingFigure 1
  • EP3516650B1 patent drawingFigure 2
  • EP3516650B1 patent drawingFigure 3

AI summary

A method for training a multi-language speech recognition network includes providing utterance datasets corresponding to predetermined languages, inserting language identification (ID) labels into the utterance datasets, wherein each of the utterance datasets is labelled by each of the language ID labels, concatenating the labeled utterance datasets, generating initial network parameters from the utterance datasets, selecting the initial network parameters according to a predetermined sequence, and training, iteratively, an end-to-end network with a series of the selected initial network parameters and the concatenated labeled utterance datasets until a training result reaches a threshold.