Contact Center Transcript Normalization with Finite-State and LLM Fallback

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Speech recognition systems in contact centers produce transcripts that are not human-friendly due to the inclusion of dates, numbers, and addresses in their exact spoken form, making them less readable compared to human-generated transcripts.

Innovation Solution

Perform named entity recognition on contact center communications to identify entities, normalize them using weighted finite-state transducers, and for entities that cannot be normalized, utilize a large language model with a confidence threshold to ensure accurate conversion.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If speech recognition systems transcribe conversations exactly as spoken, then transcription accuracy is improved, but readability deteriorates

Engineering Contradiction:
Improvetranscription accuracyVSAvoidreadability
Core Design Contradiction:
Measurement precisionVSEase of operation

Solution Approach 1:

The patent segments the transcription processing into two distinct stages: first, exact transcription is performed to capture all spoken content accurately; second, a separate normalization stage processes the transcript to convert numerical entities (dates, times, amounts) into readable formats. This segmentation allows both high transcription accuracy and improved readability to coexist without compromising either objective.

Inventive Principle:
Principle #1Segmentation

2Ease of operation

If human-generated transcripts are normalized with numerical entities in reader-friendly format, then readability is improved, but processing time increases

Engineering Contradiction:
ImprovereadabilityVSAvoidprocessing time
Core Design Contradiction:
Ease of operationVSLoss of time

Solution Approach 1:

The patent implements preliminary action by pre-compiling comprehensive lists of numerical entity patterns (dates, times, currencies, quantities) and their corresponding readable format transformations before the actual transcription normalization process. During transcript processing, the system simply matches patterns against these pre-prepared lists rather than performing complex analysis, dramatically reducing processing time while maintaining high readability.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If automatic transcripts are generated without normalization, then processing speed is improved, but readability deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoidreadability
Core Design Contradiction:
ProductivityVSEase of operation

Solution Approach 1:

The patent applies local quality by selectively normalizing only the portions of the transcript containing numerical entities (dates, times, amounts, quantities) while leaving the rest of the text in its original transcribed form. This targeted approach ensures that readability is improved precisely where needed (in numerical expressions) without adding unnecessary processing overhead to the entire transcript, thereby maintaining high processing speed while enhancing readability where it matters most.

Inventive Principle:
Principle #3Local quality

Data Source

PatentUS12353831B2Inverse text normalization of contact center communications
Publication Date: 2025.07.08 GENESYS CLOUD SERVICES INC
  • US12353831B2 patent drawing
  • US12353831B2 patent drawing
  • US12353831B2 patent drawing

AI summary

A method for inverse text normalization of contact center communications according to an embodiment includes performing named entity recognition on text from a contact center communication to identify one or more entities in the text, normalizing each of the identified one or more entities in the text using weighted finite-state transducers, and normalizing at least one entity identified in the text using a large language model in response to determining that the at least one entity identified in the text was unable to be normalized using the weighted finite-state transducers.