Speech Recognition System Using CTC and Attention Alignment

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current speech recognition systems face challenges in achieving high recognition accuracy, particularly for minor languages, due to the need for deep linguistic knowledge and the flexibility of temporal attention mechanisms, which can result in irrelevant alignments and increased inaccuracy.

Innovation Solution

The integration of Connectionist Temporal Classification (CTC) and attention-based probabilities within a deep learning framework to constrain alignments and improve recognition accuracy by combining scores from both mechanisms.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If temporal attention mechanism is used for alignment, then flexibility in alignment is improved, but recognition accuracy deteriorates due to irrelevant alignments and non-sequential mappings

Engineering Contradiction:
Improvealignment flexibilityVSAvoidrecognition accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent combines two different alignment approaches (attention mechanism and CTC alignment) into a unified framework. The attention-based alignment provides flexibility while CTC-based alignment ensures sequential consistency. By merging these two approaches and selecting the best alignment based on confidence scores, the system achieves both flexibility and accuracy in speech recognition.

Inventive Principle:
Principle #5Merging (Combining)

2Productivity

If factorized architecture with separate modules is used, then model optimization is improved, but system complexity increases and local optima are reached

Engineering Contradiction:
Improvemodel optimization efficiencyVSAvoidsystem architecture complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent merges previously separate modules (acoustic model, alignment mechanism, and decoder) into an integrated end-to-end speech recognition system. By combining attention-based alignment and CTC-based alignment within a single framework and using a unified loss function, the system reduces architectural complexity while maintaining optimization efficiency and avoiding local optima.

Inventive Principle:
Principle #5Merging (Combining)

3Measurement precision

If deep linguistic knowledge is required for system development, then recognition accuracy is improved, but ease of manufacture deteriorates for non-expert developers

Engineering Contradiction:
Improverecognition accuracyVSAvoidsystem development ease
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

The patent implements self-service by enabling the speech recognition system to automatically learn alignment patterns and linguistic features from raw speech data without requiring manual linguistic expertise. The attention mechanism and CTC framework allow the system to autonomously discover optimal alignments and phoneme sequences, making development accessible to non-experts while maintaining high accuracy.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3417451B1Speech recognition system and method for speech recognition
Publication Date: 2019.07.03 MITSUBISHI ELECTRIC CORP
  • EP3417451B1 patent drawingFigure 1
  • EP3417451B1 patent drawingFigure 2
  • EP3417451B1 patent drawingFigure 3

AI summary

A speech recognition system includes an input device to receive voice sounds, one or more processors, and one or more storage devices storing parameters and program modules including instructions executable by the one or more processors. The instructions includes extracting an accoustic feature sequence from audio waveform data converted from the voice sounds encoding the acoustic feature sequence into a hidden vector sequence (HVS) using an encoder network having encoder network parameters, predicting first output label sequence probabilities (OLSP) by feeding the (HVS) to a decoder network having decoder network parameters, predicting second (OLSP) by a connectionist temporal classification (CTC) module using CTC network parameters and the (HVS) from the encoder network, and searching, using a label sequence search module, for an output label sequence having a highest sequence probability by combining the first and second (OLSP) provided from the decoder network and the CTC module.