Speech Recognition Noise Suppression via Preset Sequence

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

End-to-end artificial neural network-based speech recognition systems often output unexpected text in noise intervals without actual user utterance, leading to forced speech recognition and failure to distinguish between noise and speech.

Innovation Solution

Incorporating a preset special sequence, such as 'hi', at the front end of the input sequence, which is recognized by the neural network to prevent noise output and focus on the main speaker's utterance through autoregressive decoding, where the previous output affects the determination of subsequent tokens.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If end-to-end artificial neural network-based speech recognition is used, then recognition speed and automation are improved, but noise output and forced speech recognition occur

Engineering Contradiction:

ImproveautomationVSAvoidnoise output

Core Design Contradiction:

Extent of automationVSReliability

Solution Approach 1:

The patent applies preliminary action by adding a preset special sequence (e.g., 'hi') to the front end of the input sequence before processing. This special sequence is recognized first by the neural network, establishing a baseline state that prevents the network from generating unexpected text during subsequent decoding of the actual speech input, thereby resolving the noise output issue while maintaining automation.

Inventive Principle:

Principle #10Preliminary action

2Speed

If real-time decoding is performed phoneme-by-phoneme, then processing speed is improved, but unexpected text is output in noise intervals

Engineering Contradiction:

Improveprocessing speedVSAvoidrecognition accuracy

Core Design Contradiction:

SpeedVSManufacturing precision

Solution Approach 1:

The preset special sequence is processed first before the actual speech input, establishing a correct initial state. This preliminary recognition prevents the phoneme-by-phoneme decoding process from generating unexpected text when encountering noise intervals, ensuring high recognition accuracy while maintaining real-time processing speed.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The preset special sequence acts as an intermediary element between the input audio signal and the final recognition output. It mediates the decoding process by providing a stable starting point that guides the neural network through noise intervals, preventing erroneous text generation while preserving processing efficiency.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Productivity

If forced speech recognition is applied, then continuous output is achieved, but noise is mistakenly recognized as speech

Engineering Contradiction:

Improvecontinuous outputVSAvoidnoise distinction

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The special sequence is recognized first to establish a baseline state, allowing the system to distinguish between actual speech and noise during subsequent processing. This preliminary action enables continuous output while improving noise distinction capability, as the network knows what to expect versus what constitutes actual speech content.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12073825B2Method and apparatus for speech recognition

Publication Date: 2024.08.27 SAMSUNG ELECTRONICS CO LTD

AI summary

A speech recognition method includes adding a preset special sequence to a front end of an input sequence that corresponds to an input utterance of a speaker, recognizing the preset special sequence and the input sequence, and recognizing the input sequence based on the preset special sequence and a speech recognition result obtained by recognizing the preset special sequence and the input sequence.