Language Identification Classifier Using Pre-Trained STT Encoder Features

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing speech-to-text systems face challenges when dealing with unknown or mismatched languages, leading to incorrect or no transcription, and require significant computational resources when handling multiple languages simultaneously.

Innovation Solution

A system that includes a processor receiving encoded audio from a pre-trained speech-to-text model and further trains a language identification classifier to detect the language of the audio using labeled training samples, allowing for efficient and accurate text transcription across multiple languages.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a language classification tool is applied to identify the language of speech before transcription, then transcription accuracy is improved, but latency increases significantly as the LID classifier requires several seconds of speech to identify the language correctly

Engineering Contradiction:

Improvetranscription accuracyVSAvoidlatency

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary language identification using the encoder features before full transcription, but optimizes this by using the pre-trained STT encoder's features rather than a separate LID classifier, reducing the preliminary processing time while still achieving accurate language detection to guide the transcription process

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent merges the language identification function with the existing speech-to-text encoder by training the encoder to output features that are directly useful for both STT and LID tasks, eliminating the need for a separate LID classifier and reducing overall system latency

Inventive Principle:

Principle #5Merging (Combining)

2Adaptability or versatility

If multiple STT modules with different languages are run in parallel to handle multiple languages simultaneously, then language coverage and transcription accuracy are improved, but computational resources increase significantly

Engineering Contradiction:

Improvelanguage coverageVSAvoidcomputational resources

Core Design Contradiction:

Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The patent creates a universal STT encoder that can handle multiple languages through a single model by training it on multi-lingual data and using language identification to route to appropriate language models, eliminating the need for multiple parallel STT modules while maintaining multi-language support

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The system introduces an intermediary language identification classifier that receives encoded audio features and determines the language, then routes the processing to the appropriate language-specific STT module, avoiding the need to run all language modules in parallel

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS12340794B2Language identification classifier trained using encoded audio from encoder of pre-trained speech-to-text system

Publication Date: 2025.06.24 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

An example system includes a processor to receive encoded audio from an encoder of a pre-trained speech-to-text (STT) model. The processor is to further train a language identification (LID) classifier to detect a language of the encoded audio using training samples labeled by language.