Speech Decoding Network Construction for Digit Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current digit speech recognition methods, whether isolated word recognition or universal continuous speech recognition, face challenges in achieving high accuracy due to the complexity of acoustic and language models, leading to misrecognition and slow recognition speeds, especially in noisy environments.

Innovation Solution

A method for constructing a speech decoding network that involves acquiring training data, performing acoustic feature extraction, and progressive training starting from a mono-phoneme acoustic model to obtain a suitable acoustic model, while also constructing a language model to improve recognition accuracy and adaptability to continuous digit speech.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If universal continuous speech recognition technology is used to recognize digits, then the system can recognize both digits and other language content, but the complexity of acoustic and language models increases leading to misrecognition and slow recognition speeds

Engineering Contradiction:
Improverecognition scopeVSAvoidmodel complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments the speech recognition system into two distinct parts: a digit recognition network specifically optimized for digit recognition tasks, and a general language model for other language content. This segmentation allows the digit recognition component to be simpler and more accurate while maintaining overall system versatility through the general language model.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements dynamic routing that determines whether to use the digit recognition network or the general language model based on the input characteristics. This dynamic approach allows the system to adapt its processing path, using the simpler digit network when appropriate and the general model when needed, thereby reducing overall complexity while maintaining versatility.

Inventive Principle:
Principle #15Dynamics

2Device complexity

If conventional HMM based only on spectral feature is used, then the model is simple, but the recognition performance is not high due to highly confusable syllables

Engineering Contradiction:
Improvemodel simplicityVSAvoidrecognition accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent creates a composite acoustic model that combines multiple feature types (spectral features, energy features, pitch contour features) rather than relying on spectral features alone. This composite approach leverages the complementary strengths of different feature types to improve recognition accuracy while maintaining reasonable model complexity through efficient feature fusion.

Inventive Principle:
Principle #40Composite materials

Solution Approach 2:

The patent changes the parameter representation by using different feature extraction parameters for different digits (e.g., emphasizing energy features for certain digits, pitch contour for others). This parameter adaptation allows the model to achieve high accuracy for highly confusable syllables by adjusting which features are most relevant for each specific digit recognition task.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If multiple feature types are used to train single-stream HMM with different emphases, then recognition accuracy improves, but the training complexity and computational load increase

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

Solution Approach 1:

The patent segments the feature processing into separate streams that are processed independently and then combined. Each feature type (spectral, energy, pitch) is processed through its own optimized pathway, which simplifies the training process compared to processing all features uniformly. This segmentation reduces training complexity while maintaining the ability to use multiple feature types for high accuracy.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3438973B1Method and apparatus for constructing speech decoding network in digital speech recognition, and storage medium
Publication Date: 2020.07.15 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3438973B1 patent drawingFigure 1~2
  • EP3438973B1 patent drawingFigure 3~4
  • EP3438973B1 patent drawingFigure 5

AI summary

A method for constructing a speech decoding network in digit speech recognition. The method comprises acquiring training data obtained by means of digit speech recording, wherein the training data comprises a plurality of speech segments, with each speech segment containing pieces of speech of a plurality of digits (110); performing acoustic feature extraction on the training data to obtain a feature sequence corresponding to each speech segment (130); according to the feature sequence and a phone corresponding to a digit in the training data, performing, starting with a mono-phone acoustic model, progressive training to obtain an acoustic model (150); and acquiring a language model, and constructing, by means of the language model and the acoustic model obtained through training, a speech decoding network (170).