End-to-End ASR Encoder-Decoder Split for Speech Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current automatic speech recognition (ASR) systems face challenges in accurately recognizing user speech due to limitations in processing and decoding speech inputs across devices and servers, particularly in integrating end-to-end ASR models effectively.

Innovation Solution

A system and method where a device encodes user speech using an end-to-end ASR model and transmits the output to a server, which decodes it using corresponding decoders associated with specific domains to provide an accurate text string, enabling improved speech recognition and natural language understanding.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If speech recognition is performed entirely on a device using a pre-trained ASR model, then processing speed is improved, but recognition accuracy deteriorates due to limited computational resources and model size

Engineering Contradiction:
Improveprocessing speedVSAvoidrecognition accuracy
Core Design Contradiction:
SpeedVSMeasurement precision

Solution Approach 1:

The ASR model is segmented into encoder and decoder components, with the encoder running on the device and the decoder running on the server. This allows the device to perform initial processing quickly while the server handles the computationally intensive decoding to achieve high accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The encoder output from the device acts as an intermediary representation that is transmitted to the server. This intermediate form contains essential speech features that can be efficiently processed by the server's decoder to achieve accurate recognition without requiring the entire model to run on the device.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If a complete end-to-end ASR model is deployed on the server, then recognition accuracy is improved, but device-server communication overhead and processing time increase

Engineering Contradiction:
Improverecognition accuracyVSAvoidcommunication overhead
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The encoder processes the speech input on the device before transmission, performing preliminary feature extraction and encoding. This preliminary action reduces the amount of data that needs to be transmitted to the server and prepares it in a format optimized for decoding, thereby reducing communication overhead and total processing time.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If domain-specific decoders are used on the server, then speech recognition accuracy for specific domains is improved, but system complexity increases

Engineering Contradiction:
Improvedomain-specific recognition accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

Different domain-specific decoders are deployed on the server to handle different types of speech inputs (e.g., commands, conversations, specialized terminology). Each decoder is optimized for its specific domain, providing high accuracy for that domain while the overall system maintains modularity to manage complexity.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3980991B1System and method for recognizing user's speech
Publication Date: 2024.01.03 SAMSUNG ELECTRONICS CO LTD
  • EP3980991B1 patent drawingFigure 1
  • EP3980991B1 patent drawingFigure 2
  • EP3980991B1 patent drawingFigure 3

AI summary

Provided is a system and method for recognizing a user's speech. A method, performed by a server, of providing a text string for a speech signal input to a device includes: receiving, from the device, an encoder output value derived from an encoder of an end-to-end automatic speech recognition (ASR) model included in the device; identifying a domain corresponding to the received encoder output value; selecting a decoder corresponding to the identified domain from among a plurality of decoders of an end-to-end ASR model included in the server; obtaining a text string from the received encoder output value using the selected decoder; and providing the obtained text string to the device.