End-to-End Speech Recognition Model Using Attention Encoding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Hybrid speech recognition systems have low accuracy due to the poor coordination between multiple modules such as the acoustic model, pronunciation dictionary, and language model, leading to inefficient speech recognition in noisy or inconvenient environments.

Innovation Solution

A cloud-based end-to-end speech recognition system using a TL-AED model with attention encoding and sequential decoding, which simplifies the architecture and improves recognition accuracy by reducing the number of parameters and utilizing parallel computation effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If hybrid speech recognition systems use multiple modules (acoustic model, pronunciation dictionary, language model), then the system can process speech recognition, but the coordination between modules is poor and recognition accuracy is low

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidsystem architecture complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges the acoustic model, pronunciation dictionary, and language model into a unified end-to-end neural network architecture. This integration eliminates the coordination problems between separate modules while maintaining the functional capabilities of speech recognition, thereby improving accuracy and reducing system complexity.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The end-to-end neural network is designed to perform multiple functions simultaneously: acoustic feature extraction, phoneme recognition, and language modeling within a single unified framework. This multi-functional design improves coordination and reduces the complexity associated with managing multiple specialized modules.

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

2Speed

If the speech recognition system processes speech in real-time, then the user experience is improved, but the computational complexity increases

Engineering Contradiction:
Improvereal-time processing speedVSAvoidcomputational complexity
Core Design Contradiction:
SpeedVSDevice complexity

Solution Approach 1:

The patent segments the speech signal into frames and processes them sequentially through the neural network. This segmentation allows for efficient real-time processing by breaking down the continuous speech stream into manageable units that can be processed in parallel where possible, reducing overall computational complexity while maintaining real-time performance.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary feature extraction and preprocessing of speech signals before main recognition processing. This preliminary action prepares the data in an optimized format that reduces the computational burden during real-time recognition, enabling faster processing without sacrificing accuracy.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If the neural network uses high numerical precision for weights and activations, then the model performance is maintained, but the computational resources and memory usage increase

Engineering Contradiction:
Improvemodel performanceVSAvoidcomputational resource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent employs mixed precision arithmetic, using different numerical precisions for different parts of the neural network computation. Critical operations that require high accuracy use higher precision, while less sensitive operations use lower precision, thereby maintaining model performance while significantly reducing memory usage and computational resource consumption.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The system uses lower precision data types (such as float16 or bfloat16) for intermediate computations and activations that can tolerate some precision loss. These lower-precision representations consume less memory and computational resources, and the model is trained to be robust to this reduced precision, effectively using 'cheaper' computational objects where appropriate.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Data Source

PatentEP4113507B1Speech recognition method and device
Publication Date: 2024.11.13 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP4113507B1 patent drawingFigure 1
  • EP4113507B1 patent drawingFigure 2
  • EP4113507B1 patent drawingFigure 3

AI summary

A speech recognition method and apparatus, a device, and a storage medium, relating to the field of speech recognition. Speech recognition is performed using attention coding and time series decoding methods, and attention coding is performed on a speech feature matrix, such that the parallel computing of a GPU can be efficiently utilized. Decoding coded vectors according to the positions of the coded vectors in a coded matrix can reduce the amount of parameters, and the use of memory features of a speech recognition model to perform prediction according to time-series related coded vectors can improve speech recognition accuracy.