Utterance Time Estimation via Attention Weight Matrices

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition methods, particularly those based on hidden Markov models (HMM) and transformer architectures, struggle to accurately estimate utterance times for output sequences without analyzing pronunciation from audio data, leading to inefficiencies and inaccuracies in speech recognition systems.

Innovation Solution

A processor-implemented utterance time estimation method using an attention-based sequence-to-sequence model, which determines a plurality of attention weight matrices, selects a suitable attention weight matrix through monotonic regression analysis, and estimates utterance times based on high attention weights and diffusion corrections.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If HMM-based speech recognition method is used to analyze pronunciation from audio data, then speech recognition can be performed, but utterance time estimation accuracy deteriorates and processing efficiency decreases

Engineering Contradiction:
Improveutterance time estimation accuracyVSAvoidprocessing efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent replaces the traditional HMM-based mechanical analysis approach with an attention-based neural network approach. The attention-based sequence-to-sequence model directly estimates utterance times from audio data without requiring explicit pronunciation analysis, substituting the complex multi-step mechanical process with a more efficient neural network-based solution that achieves both higher accuracy and better processing efficiency

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the fundamental parameters of the speech recognition system by using attention weight matrices instead of traditional HMM pronunciation analysis parameters. By transforming the problem into an attention-based estimation task, the system can directly output utterance time predictions without the intermediate pronunciation analysis step, thereby improving both accuracy and efficiency

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If traditional speech recognition methods are used, then speech can be recognized, but utterance time estimation requires additional analysis steps increasing device complexity

Engineering Contradiction:
Improveutterance time estimation accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges the speech recognition and utterance time estimation functions into a single integrated attention-based sequence-to-sequence model. Instead of using separate HMM-based pronunciation analysis followed by time estimation, the model simultaneously performs both tasks, reducing system complexity while maintaining or improving estimation accuracy through the unified attention mechanism

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The attention-based sequence-to-sequence model serves multiple functions: it performs speech recognition, extracts pronunciation information, and estimates utterance times all within a single framework. This multi-functional approach eliminates the need for separate specialized components, thereby reducing overall device complexity while achieving accurate utterance time estimation

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

3Productivity

If attention-based sequence-to-sequence model is used to directly estimate utterance time, then processing efficiency improves, but model complexity increases

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidmodel complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent segments the attention-based model into manageable components: an encoder that processes audio input, a decoder that generates output sequences, and an attention mechanism that weights important temporal regions. This segmentation allows the complex model to be implemented through standard neural network building blocks, improving processing efficiency while keeping the complexity manageable through modular architecture

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3910625B1Method and apparatus for utterance time estimation
Publication Date: 2025.06.11 SAMSUNG ELECTRONICS CO LTD
  • EP3910625B1 patent drawingFigure 1
  • EP3910625B1 patent drawingFigure 2
  • EP3910625B1 patent drawingFigure 3

AI summary

The application concerns a processor-implemented utterance time estimation method and utterance time estimation apparatus, comprising: determining a plurality of attention weight matrices using an attention-based sequence-to-sequence model; selecting an attention weight matrix from the plurality of attention weight matrices; and estimating an utterance time corresponding to an output sequence based on the selected attention weight matrix. A method may also include encoding an input sequence including input frames, estimating output tokens of an output sequence corresponding to the input sequence, based on the encoded feature, determining attention weight matrices including attention weights, each attention weight corresponding to one of the input frames and one of the output tokens, and estimating an utterance time corresponding to an output sequence based on the attention weight matrices.