Speech Recognition Blank Suppression for Fewer Deletion Errors

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The introduction of blank output in phoneme recognition processes using RNN-T models increases error rates, particularly deletion errors, affecting the accuracy of speech recognition in certain scenarios.

Innovation Solution

A speech recognition method that performs suppression adjustment on phoneme recognition results to reduce the probability of blank outputs, using a decoding map with threshold conditions and influence parameters to improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If blank output is introduced in phoneme recognition process using RNN-T model, then the model can handle silent frames better, but the error rate increases particularly deletion errors

Engineering Contradiction:
Improvehandling of silent framesVSAvoidrecognition accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent applies parameter changes by dynamically adjusting the probability threshold for blank output based on multiple influence parameters including ambient sound intensity, number of recognition failures, and user settings. This allows the system to adaptively control when to accept blank outputs versus when to enforce phoneme recognition, resolving the contradiction between handling silent frames and maintaining recognition accuracy.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The system implements dynamics by making the blank output threshold dynamic rather than fixed. The threshold varies based on real-time conditions such as ambient noise levels and recognition history, allowing the system to flexibly balance between accepting blank outputs and preventing deletion errors depending on the current operational context.

Inventive Principle:
Principle #15Dynamics

2Measurement precision

If suppression adjustment is performed to reduce blank output probability, then deletion errors decrease, but the system complexity increases

Engineering Contradiction:
Improverecognition accuracyVSAvoidprocessing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by performing suppression adjustment on phoneme recognition results before they are used in the decoding process. By pre-adjusting the probability distributions and filtering out unlikely blank outputs in advance, the system reduces deletion errors without significantly increasing overall processing complexity, as the adjustment operates on already-computed recognition results.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If dynamic threshold adjustment based on multiple parameters is implemented, then recognition accuracy improves, but computational overhead increases

Engineering Contradiction:
Improverecognition accuracyVSAvoidCPU usage
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system applies partial action by selectively applying threshold adjustment only when necessary, based on the evaluated influence parameters. Rather than continuously adjusting thresholds for all frames, the system determines the appropriate threshold behavior based on ambient conditions and recognition history, applying adjustments only when they are likely to improve accuracy, thus reducing unnecessary computational overhead.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4191576B1Speech recognition method, computer device, and storage medium
Publication Date: 2025.12.31 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP4191576B1 patent drawingFigure 1~2
  • EP4191576B1 patent drawingFigure 3~4
  • EP4191576B1 patent drawingFigure 5~6

AI summary

A speech recognition method and apparatus, a computer device, and a storage medium. The method comprises: performing phoneme recognition on a speech signal to obtain a phoneme recognition result corresponding to each speech frame in the speech signal (21); performing suppression adjustment on a probability of blank outputs in the phoneme recognition result corresponding to each speech frame, so as to reduce a ratio of the probability of the blank outputs in the phoneme recognition result to a probability of each phoneme (22); and inputting the adjusted phoneme recognition result corresponding to each speech frame into a decoding map to obtain a recognition text sequence corresponding to the speech signal (23). The method can improve the recognition accuracy of a model in a speech recognition scenario of the field of artificial intelligence.