Two-Stage Speech Recognition Using Word-Level Audio Features

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition methods face challenges with inconsistent feature representation lengths and redundant features due to varying speech speeds and tones, leading to reduced accuracy and computational efficiency.

Innovation Solution

A method that involves obtaining a first speech feature, decoding it to get a preliminary recognition result, extracting word-level audio features, and then decoding these features using a second decoder to achieve uniform feature representation, improving precision and efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a single decoder is used to decode speech features, then the device complexity is low, but the recognition accuracy is insufficient

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

Solution Approach 1:

The patent divides the decoding process into multiple independent decoders, where each decoder is responsible for decoding a specific subset of speech segment features. This segmentation allows each decoder to specialize in particular acoustic patterns or phonetic categories, thereby improving overall recognition accuracy while maintaining manageable complexity through modular design

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension to the decoding process by using multiple decoders that operate in parallel on different feature subsets. This dimensional expansion from a single decoder to multiple decoders enables the system to capture more diverse speech patterns simultaneously, improving recognition accuracy without proportionally increasing overall system complexity

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If multiple decoders are used to decode different speech segment features, then the recognition accuracy is improved, but the device complexity increases

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

Solution Approach 1:

The patent segments the speech feature decoding task into multiple sub-tasks, with each decoder handling a specific subset of features. This segmentation strategy improves accuracy by allowing specialized decoding for different acoustic patterns while controlling complexity through clear task division

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent designs the multiple decoders to work within a unified framework where each decoder, though specialized, contributes to a common recognition objective. The decoders share common architectural elements and are coordinated through the same training process, providing multi-functionality that improves accuracy without linearly increasing complexity

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If speech features are decoded into word-level features, then the recognition accuracy is improved, but the training time increases

Engineering Contradiction:
Improverecognition accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary feature extraction and encoding of speech segments into a standardized feature representation before the actual decoding process. This preliminary action prepares the data in advance, allowing the multiple decoders to operate more efficiently during training and reducing overall training time while maintaining high accuracy

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the training process into independent stages where each decoder can be trained separately on its specific feature subset. This segmentation enables parallel training operations, significantly reducing total training time while maintaining the benefits of word-level feature decoding for improved accuracy

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4475119B1Speech recognition method, method for training deep learning model for speech recognition, apparatus and device
Publication Date: 2026.04.29 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP4475119B1 patent drawingFigure 1
  • EP4475119B1 patent drawingFigure 2
  • EP4475119B1 patent drawingFigure 3

AI summary

The present disclosure provides a speech recognition method, a method for training a deep learning model, an apparatus and a device, which relate to the technical field of artificial intelligence, in particular to the technical field of speech recognition and deep learning etc. The speech recognition method comprises: obtaining a first speech feature of a speech to-be-recognized, where the first speech feature comprises a plurality of speech segment features corresponding to a plurality of speech segments; decoding the first speech feature using a first decoder to obtain a plurality of first decoding results corresponding to a plurality of words, where the first decoding result indicates the first recognition result of a corresponding word; extracting a second speech feature from the first speech feature based on first a priori information, where the first a priori information comprises the plurality of first decoding results, and the second speech feature comprises a plurality of first word-level audio features corresponding to the plurality of words; and decoding the second speech feature using a second decoder to obtain a plurality of second decoding results corresponding to the plurality of words, where the second decoding result indicates the second recognition result of the corresponding word.