Event Mask Generation for Audio Signal Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing audio signal processing technologies struggle to detect nonstationary sounds with unknown spectrum shapes as audio events, leading to incomplete detection and omission of audio events.

Innovation Solution

A mask generation device and method that extracts sound pressure information from a spectrogram and applies a binarization process to generate an event mask, indicating the time period of an audio event, allowing for the detection of audio events with unknown spectrum shapes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If existing voice activity detection methods are used to distinguish voice from noise, then voice sections can be identified, but nonstationary sounds with unknown spectrum shapes cannot be detected as audio events

Engineering Contradiction:

Improvedetection capability for unknown spectrum shapesVSAvoiddetection accuracy for nonstationary sounds

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The invention changes the detection parameter from assuming a specific spectrum shape (voice) to detecting sound pressure level changes regardless of spectrum shape. By using binarization of sound pressure information extracted from spectrogram, the system can detect both stationary and nonstationary sounds with any spectrum shape as audio events, resolving the contradiction between adaptability to unknown spectra and reliability in detection accuracy.

Inventive Principle:

Principle #35Parameter changes

2Measurement precision

If temporal waveform is used to determine sound pressure, then voice detection is possible, but sounds with strong power in specific frequencies but unknown spectrum shape cannot be sufficiently detected

Engineering Contradiction:

Improvesound pressure measurement accuracyVSAvoiddetection capability for frequency-specific sounds

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The invention introduces spectrogram as an intermediary representation between the temporal waveform and the sound detection. The spectrogram provides frequency-domain information that allows extraction of sound pressure information for each frequency component. By binarizing this extracted information, the system achieves both precise sound pressure measurement and adaptability to detect frequency-specific nonstationary sounds with unknown spectrum shapes.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Reliability

If event mask is applied to spectrogram to reduce noise, then voice recognition accuracy improves, but nonstationary audio events with unknown spectrum shapes are omitted

Engineering Contradiction:

Improvevoice recognition accuracyVSAvoidomission of audio event detection

Core Design Contradiction:

ReliabilityVSLoss of information

Solution Approach 1:

Instead of applying a fixed event mask to suppress noise and potentially losing information, the invention inverts the approach by first extracting sound pressure information from the spectrogram and then applying binarization to create an adaptive event mask. This inverted approach ensures that both stationary and nonstationary sounds are detected as audio events, preventing information loss while maintaining noise reduction capabilities for voice recognition.

Inventive Principle:

Principle #13The other way round (Inversion)

Data Source

PatentUS11881200B2Mask generation device, mask generation method, and recording medium

Publication Date: 2024.01.23 NEC CORP

AI summary

An extraction unit extracts sound pressure information from a spectrogram. A binarization unit carries out binarization on the extracted sound pressure information in order to generate an event mask indicating a time period in which an audio event exists.