Speaker Vector Integration in RNN Language Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional recurrent neural network-based language models do not account for speakers, making it impossible to explicitly use information about who has spoken and who will speak next in multi-person conversations, limiting their effectiveness in speech recognition and other applications.

Innovation Solution

A language model score calculation apparatus that incorporates speaker information by converting speaker labels into speaker vectors and combining them with word history vectors to calculate prediction probabilities, allowing the model to account for speaker-specific characteristics.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If conventional recurrent neural network based language model is used, then language prediction performance is high, but speaker information cannot be explicitly utilized

Engineering Contradiction:
Improvespeaker informationVSAvoidmodel structure
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent embeds speaker identity information within the existing word history vector structure. The speaker vector is integrated into the recurrent neural network's hidden state, allowing speaker information to be nested within the same computational framework that processes word sequences, thus preserving information without significantly increasing external complexity

Inventive Principle:
Principle #7Nested doll (Nesting)

Solution Approach 2:

The patent introduces a speaker vector as an intermediary representation that bridges speaker identity and the language model. This speaker vector acts as a mediator that encodes speaker-specific characteristics and is combined with word history vectors, enabling the model to utilize speaker information without directly modifying the core recurrent neural network architecture

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If speaker information is incorporated into the language model, then word prediction sophistication is improved, but model complexity increases

Engineering Contradiction:
Improveword prediction accuracyVSAvoidmodel structure
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies local quality by maintaining different representation types for different aspects of information. Word vectors capture lexical meaning while speaker vectors capture speaker-specific characteristics. Each type of vector maintains its own properties and is processed appropriately within the network, allowing the model to handle multiple types of information with specialized representations rather than forcing a uniform structure

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent merges speaker vectors with word history vectors in the recurrent neural network to create a unified representation that incorporates both speaker identity and linguistic context. This combining is done through concatenation or addition operations that integrate the two types of information while maintaining computational efficiency, thus improving prediction accuracy without proportionally increasing complexity

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentUS12131729B2Language model score calculating apparatus, learning apparatus, language model score calculating method, learning method and program
Publication Date: 2024.10.29 NIPPON TELEGRAPH & TELEPHONE CORP
  • US12131729B2 patent drawing
  • US12131729B2 patent drawing
  • US12131729B2 patent drawing

AI summary

A language model score calculation apparatus calculates a prediction probability of a word wi as a language model score of a language model based on a recurrent neural network. The language model score calculation apparatus includes a memory; and a processor configured to execute converting a word wi-1 that is observed immediately before the word wi into a word vector ϕ(wi-1); converting a speaker label ri-1 corresponding to the word wi-1 and a speaker label ri corresponding to the word wi into a speaker vector ψ(ri-1) and a speaker vector ψ(ri), respectively; calculating a word history vector si by using the word vector ϕ(wi-1), the speaker vector ψ(ri-1), and a word history vector si-1 that is obtained when a prediction probability of the word wi-1 is calculated; and calculating a prediction probability of the word wi by using the word history vector si-1 and the speaker vector ψ(ri).