Personalized Speech Recognition via Cepstral Parameter Extraction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current speech recognition technologies face challenges in achieving accurate and personalized recognition, particularly in adapting to individual user characteristics such as accent, tone, and environment, which affects the reliability and effectiveness of speech interfaces.

Innovation Solution

A system and method that utilize a user terminal to determine personalized characteristic parameters, such as cepstral mean and variance normalization, i-vectors, and probability density functions, which are then used to adapt a general-purpose speech recognition model on a server for improved accuracy, without exposing personal information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a general-purpose speech recognition model is used, then the system can recognize speech from any user, but the recognition accuracy decreases for individual users due to variations in accent, tone, and environment

Engineering Contradiction:
Improveuser adaptabilityVSAvoidspeech recognition accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The system performs preliminary action by determining user-specific characteristic parameters (such as accent, tone, and environmental characteristics) before the actual speech recognition process. These parameters are extracted in advance and stored for later use, allowing the recognition model to be pre-adapted to individual users, thereby improving recognition accuracy without compromising system versatility

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system applies local quality by maintaining a general-purpose recognition model for overall versatility while introducing user-specific characteristic parameters that locally adapt the recognition process to individual users. This allows different parts of the recognition system to have different characteristics - the base model handles general speech recognition while user-specific parameters handle individual variations in accent, tone, and environment

Inventive Principle:
Principle #3Local quality

2Measurement precision

If user-specific parameters are collected and stored for personalization, then speech recognition accuracy improves, but user privacy and security risks increase

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidprivacy risk
Core Design Contradiction:
Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The system extracts only the necessary characteristic parameters (such as accent, tone, and environmental features) from user speech data, rather than storing complete speech recordings or personal information. This extraction approach maintains recognition accuracy by capturing essential user characteristics while minimizing privacy risks by removing unnecessary personal data

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system uses characteristic parameters as an intermediary representation of user-specific features. Instead of directly using raw speech data or personal information, the system transforms user characteristics into abstract parameters that can be used for personalization while serving as a protective layer between user privacy and the recognition system

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3444809B1Personalized speech recognition method and system
Publication Date: 2020.09.23 SAMSUNG ELECTRONICS CO LTD
  • EP3444809B1 patent drawingFigure 1
  • EP3444809B1 patent drawingFigure 2
  • EP3444809B1 patent drawingFigure 3

AI summary

A recognition method performed in a user terminal includes determining a characteristic parameter personalized to a speech of a user based on a reference speech signal input by the user; receiving, as an input, a target speech signal to be recognized from the user; and outputting a recognition result of the target speech signal, wherein the recognition result of the target speech signal is determined based on the characteristic parameter and a model for recognizing the target speech signal.