End-to-End Speech Recognition Segmentation for Delay Reduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing end-to-end speech recognition systems face challenges in achieving high recognition accuracy and reducing output delays, particularly for minor languages and online/streaming applications, due to the complexity of linguistic knowledge requirements and the need for large input sequence processing by attention-based neural networks.

Innovation Solution

A speech recognition system is developed that includes an encoder network, an alignment network (such as a CTC-based or HMM-based model), and an attention-based neural network, where the alignment network identifies relevant locations in the input sequence to constrain the attention-based network, allowing for partitioning of the input sequence into smaller subsets for sequential processing, thereby reducing output delays and improving accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If attention-based neural networks process large input sequences to assign weights to each input frame, then recognition accuracy is improved, but output delays increase

Engineering Contradiction:
Improverecognition accuracyVSAvoidoutput delays
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments the large input sequence into multiple smaller sub-sequences or frames, processing them in chunks rather than as a single large sequence. This allows the attention mechanism to operate on manageable portions of data, reducing the computational burden and output delay while maintaining recognition accuracy through progressive processing of segmented input data.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary filtering or pre-processing to identify and prioritize relevant input frames before feeding them to the attention-based network. By pre-identifying which input frames are most likely to be relevant based on initial analysis or heuristics, the system reduces the effective sequence length that requires full attention processing, thereby reducing output delays while preserving accuracy.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If attention-based networks observe entire speech utterances to assign weights to input frames, then transcription accuracy is improved, but computational complexity increases

Engineering Contradiction:
Improvetranscription accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent divides the entire speech utterance into smaller temporal segments or frames that can be processed independently or in parallel. This segmentation reduces the computational complexity of the attention mechanism by limiting the scope of interactions to local segments rather than requiring global attention across the entire utterance, while still achieving accurate transcription through hierarchical or sequential processing of segments.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements partial attention by focusing computational resources only on relevant portions of the input sequence rather than processing the entire sequence with full attention. By identifying and attending only to critical segments that contain meaningful speech information, the system achieves accurate transcription with reduced computational complexity compared to exhaustive full-sequence attention.

Inventive Principle:
Principle #16Partial or excessive action

3Ease of manufacture

If conventional ASR systems use factorized architecture with separate acoustic, lexicon, and language models, then each model can be optimized independently, but the system achieves local optima and requires deep linguistic knowledge

Engineering Contradiction:
Improvemodel optimizationVSAvoidrecognition accuracy
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent merges the previously separate acoustic model, lexicon, and language model into a unified attention-based neural network architecture. This end-to-end approach combines all functions into a single trainable system that processes acoustic input and generates transcription output directly, eliminating the need for separate model optimization and external linguistic resources while achieving superior recognition accuracy through joint learning of all components.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3948850B1System and method for end-to-end speech recognition with triggered attention
Publication Date: 2023.01.18 MITSUBISHI ELECTRIC CORP
  • EP3948850B1 patent drawingFigure 1
  • EP3948850B1 patent drawingFigure 2A
  • EP3948850B1 patent drawingFigure 2B

AI summary

A speech recognition system includes an encoder to convert an input acoustic signal into a sequence of encoder states, an alignment decoder to identify locations of encoder states in the sequence of encoder states that encode transcription outputs, a partition module to partition the sequence of encoder states into a set of partitions based on the locations of the identified encoder states, and an attention-based decoder to determine the transcription outputs for each partition of encoder states submitted to the attention-based decoder as an input. Upon receiving the acoustic signal, the system uses the encoder to produce the sequence of encoder states, partitions the sequence of encoder states into the set of partitions based on the locations of the encoder states identified by the alignment decoder, and submits the set of partitions sequentially into the attention-based decoder to produce a transcription output for each of the submitted partitions.