Dynamic Language Models for Irrelevant Speech Filtering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional speech recognition systems struggle with recognizing irrelevant speech in multi-user scenarios, leading to poor recognition effects and user experience, particularly in vehicles, where dialogues between users and in-vehicle modules interfere with intended commands.

Innovation Solution

A dynamic target language model is employed, comprising a front-end part for confirmatory information and a core part for determining relevant speech, ignoring irrelevant dialogues, and a rear-end part for detecting additional intents, using context detection and historical speech signals to improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a conventional language model is used to recognize all speech inputs, then the system can process any speech signal, but it cannot distinguish between relevant and irrelevant speech, leading to poor recognition accuracy in multi-user scenarios

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidspeech input coverage
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The language model is divided into multiple specialized models, each trained to recognize speech from a specific user or for a specific task. The system segments the speech recognition function by creating dedicated models (e.g., first language model for user 1, second language model for user 2) rather than using a single general model, thereby improving accuracy for each segment while maintaining overall system versatility.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different parts of the speech recognition system are assigned different qualities or specializations. Each language model has local quality optimized for its specific purpose (e.g., recognizing commands from a particular user), while the overall system maintains versatility through the collection of these specialized models. The system selects and applies the appropriate local model based on the speech source.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If the language model is adjusted to recognize vocabularies related to a question, then it can accurately recognize reply speech, but it also recognizes irrelevant speech from other users, causing service deviation

Engineering Contradiction:
Improvereply speech recognition accuracyVSAvoidservice accuracy
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The system dynamically selects which language model to use based on the speech source and context. Instead of a static model that always recognizes all vocabularies, the system activates the appropriate specialized model (first language model or second language model) depending on which user is speaking, making the recognition system adaptive and reliable in multi-user scenarios.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system incorporates feedback mechanisms to determine whether detected speech is relevant or irrelevant. By monitoring speech characteristics, user identification, and context, the system provides feedback to the language model selection process, ensuring that the appropriate model is activated only when relevant speech is detected, thereby maintaining service accuracy.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3965101B1Speech recognition method, apparatus and device, and computer-readable storage medium
Publication Date: 2025.08.27 YINWANG INTELLIGENT TECHNOLOGIES CO LTD
  • EP3965101B1 patent drawingFigure 1
  • EP3965101B1 patent drawingFigure 2
  • EP3965101B1 patent drawingFigure 3

AI summary

A speech recognition method, apparatus, and device, and a computer-readable storage medium are provided, pertain to the field of artificial intelligence technologies, and are particularly applicable to man-machine interaction in an automobile or an electric vehicle. The method includes: (201) obtaining or generating a dynamic target language model based on reply information of a first intent, where the dynamic target language model includes a front-end part and a core part; (202) obtaining a speech signal, and parsing the speech signal to generate a key word; and (203) invoking the dynamic target language model to determine a second intent and service content, where the front-end part of the dynamic target language model parses out the second intent based on the key word, and the core part of the dynamic target language model parses out the service content based on the key word. According to the speech recognition method, irrelevant speeches can be ignored, and discontinuous multi-intent speeches can be recognized, to prevent provided service content from deviating from a user requirement, and achieve a good recognition effect.