Concatenated Audio Training Data for Domain-Adaptive E2E ASR

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

End-to-end automatic speech recognition (E2E ASR) systems face challenges in adapting to new domains due to the high cost and time required for gathering paired audio and text data, and they are not as robust in diverse environments compared to hybrid approaches.

Innovation Solution

An audio concatenation technique generates paired audio and text data by combining audio segments from existing training data to create augmented training data, which can be used to train E2E models like RNN-T and AED, improving robustness and reducing the need for new data collection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If paired audio and text data is collected for training E2E models, then model training accuracy is improved, but resource cost and time consumption increase

Engineering Contradiction:
Improvemodel training accuracyVSAvoiddata collection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary actions by collecting and storing audio segments from diverse sources (different speakers, accents, and acoustic environments) in advance. These pre-collected audio segments are then rapidly assembled using text-to-speech synthesis to create training pairs, eliminating the need for time-consuming manual data collection for each training set.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system uses text-to-speech synthesis to generate synthetic audio copies of text data. By converting text into synthesized speech audio, the system creates artificial training pairs without requiring actual recorded human speech, thus reducing data collection time while maintaining training effectiveness.

Inventive Principle:
Principle #26Copying

2Reliability

If diverse audio data is collected for robustness, then model performance in varied environments is improved, but data collection complexity increases

Engineering Contradiction:
Improvemodel robustnessVSAvoiddata collection complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system employs a universal text-to-speech synthesis engine that can generate audio data across multiple conditions (different speakers, accents, noise environments) from a single text input. This multi-functional approach replaces the need for separate data collection processes for each condition, simplifying the overall data collection complexity while maintaining model robustness.

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

Solution Approach 2:

The system changes acoustic parameters (noise levels, speaker characteristics, accents) through signal processing and synthesis techniques. By manipulating these parameters computationally rather than collecting data under each condition separately, the system achieves diverse training data with reduced collection complexity.

Inventive Principle:
Principle #35Parameter changes

3Manufacturing precision

If manual data annotation is performed, then data quality is improved, but processing time increases

Engineering Contradiction:
Improvedata qualityVSAvoiddata processing speed
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The system uses automated text-to-speech synthesis to generate audio data with inherent text-transcription pairs, eliminating the need for manual annotation. The synthesized audio automatically comes with its corresponding text transcript, providing high-quality training data at scale without human intervention in the annotation process.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4264597B1Augmented training data for end-to-end models
Publication Date: 2025.12.24 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP4264597B1 patent drawingFigure 1
  • EP4264597B1 patent drawingFigure 2
  • EP4264597B1 patent drawingFigure 3

AI summary

A computer system is provided that includes a processor configured to store a set of audio training data that includes a plurality of audio segments and metadata indicating a word or phrase associated with each audio segment. For a target training statement of a set of structured text data, the processor is configured to generate a concatenated audio signal that matches a word content of a target training statement by comparing the words or phrases of a plurality of text segments of the target training statement to respective words or phrases of audio segments of the stored set of audio training data, selecting a plurality of audio segments from the set of audio training data based on a match in the words or phrases between the plurality of text segments of the target training statement and the selected plurality of audio segments, and concatenating the selected plurality of audio segments.