Speech Recognition With Bias Encoding for Low-Frequency Hot Words

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional speech recognition models have poor recognition accuracy for low-frequency words, particularly hot words, due to over-confidence in high-frequency words and inadequate handling of hot words during the decoding process.

Innovation Solution

A modified speech recognition model structure that includes a Bias encoder to encode hot words, utilizing an audio encoding feature and a hot word encoding feature processed by a joint attention mechanism to obtain accurate audio-related and hot word-related features for improved decoding and classification.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a conventional speech recognition model processes speech data, then the recognition speed is fast, but the recognition accuracy for hot words is poor

Engineering Contradiction:
Improverecognition accuracyVSAvoidmodel structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The speech recognition model is segmented into distinct functional modules: an audio encoding module for processing speech audio data, a hot word encoding module for encoding hot word text data, and a joint attention module for fusing features. This segmentation allows each module to specialize in specific tasks, improving hot word recognition accuracy without overwhelming the overall system complexity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

A joint attention mechanism is introduced as an intermediary between the audio encoding module and hot word encoding module. This intermediary processes both audio and text features simultaneously, enabling the model to attend to relevant hot word information in the speech data while maintaining the原有的 fast recognition speed through efficient feature fusion.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If the speech recognition model uses a large amount of training data, then the recognition effect for high-frequency words is good, but the recognition effect for low-frequency hot words deteriorates

Engineering Contradiction:
Improverecognition accuracy for hot wordsVSAvoidmodel over-confidence
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The hot word encoding module performs preliminary encoding of hot word text data before the main recognition process. By pre-processing and embedding hot word information separately, the model prepares relevant features in advance, enabling it to accurately recognize hot words during decoding without being biased by the over-confidence issue that affects standard training approaches.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The model changes the parameter representation by introducing separate encoding pathways for hot words. Instead of relying solely on standard training data distributions that create over-confidence, the model transforms the input parameters through dedicated hot word encoding, allowing it to adjust its confidence levels appropriately for low-frequency hot words while maintaining accuracy.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4156176B1Speech recognition method, apparatus and device, and storage medium
Publication Date: 2025.09.10 IFLYTEK CO LTD
  • EP4156176B1 patent drawingFigure 1~2
  • EP4156176B1 patent drawingFigure 3
  • EP4156176B1 patent drawingFigure 4

AI summary

A speech recognition method, apparatus and device, and a storage medium. The method comprises: acquiring a configured hot word library; during the process of recognizing speech to be recognized, determining, on the basis of said speech and the hot word library, an audio-related characteristic needed at the current decoding moment; since hot word information is used during the process of determining the audio-related characteristic, if a speech clip at the current decoding moment includes a certain hot word, the determined audio-related characteristic being capable of including complete audio information corresponding to the hot word; further determining, on the basis of the audio-related characteristic and from the hot word library, a hot-word-related characteristic needed at the current decoding moment, wherein the hot-word-related characteristic can accurately indicate whether the speech clip at the current decoding moment includes a hot word and indicate which hot word is specifically included; and finally, determining a recognition result of said speech at the current decoding moment on the basis of the audio-related characteristic and the hot-word-related characteristic, wherein the recognition result for recognition of the hot word is more accurate.