ASR Intent Classification Using Embedding Alignment

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing natural language processing systems face challenges in accurately determining intent from automatic speech recognition (ASR) transcriptions due to errors, which affect downstream analytics, and are constrained by regulatory restrictions on processing speech data.

Innovation Solution

A framework that trains an encoder using both ground truth and ASR transcriptions, employing triplet loss and Kullback-Leibler divergence loss to align and correct embeddings, thereby improving intent classification accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If ASR transcriptions are used for intent classification, then processing speed and compliance with regulatory restrictions are improved, but classification accuracy deteriorates due to transcription errors

Engineering Contradiction:

Improveintent classification accuracyVSAvoidtranscription errors

Core Design Contradiction:

ReliabilityVSLoss of information

Solution Approach 1:

The patent introduces an intermediary correction mechanism that uses the ASR transcription as a base and applies error correction through multiple techniques (confidence threshold filtering, n-gram analysis, and contextual replacement) to produce a corrected transcription that maintains the benefits of ASR processing while reducing its errors for improved intent classification

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The system implements feedback loops where the intent classification results and confidence scores are used to identify and correct errors in the ASR transcription. The corrected transcription is then fed back into the classification process, creating an iterative improvement cycle that enhances accuracy while maintaining processing efficiency

Inventive Principle:

Principle #23Feedback

2Measurement precision

If manual transcription is used instead of ASR, then transcription accuracy is improved, but processing time and cost increase

Engineering Contradiction:

Improvetranscription accuracyVSAvoidprocessing speed

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent applies partial correction action by only correcting transcriptions that fall below a confidence threshold or exhibit specific error patterns, rather than applying manual-level correction to all transcriptions. This selective approach maintains high processing speed while improving accuracy for the most problematic cases

Inventive Principle:

Principle #16Partial or excessive action

Solution Approach 2:

The system dynamically adjusts correction parameters such as confidence thresholds, n-gram window sizes, and replacement strategies based on the characteristics of the ASR output and the specific domain context, optimizing the balance between processing speed and transcription accuracy for different scenarios

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12609113B2Natural language processing systems and methods for intent classification of speech transcription

Publication Date: 2026.04.21 JPMORGAN CHASE BANK NA

AI summary

Aspects of the subject disclosure may include, for example, generating a natural language processing model by training an automatic speech recognition (ASR) encoder with manual transcription. The training is performed by correcting and adjusting relevant factors of the ASR encoder based on determined triplet loss, classification loss and Kullback-Leibler divergence loss. In response to an ASR utterance, the trained natural language processing model generates a predicted intent associated with the ASR utterance with improved accuracy. Other embodiments are disclosed.