Multi-Speaker Tag Estimation With Speaker-Aware Sequence Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional scene tagging techniques for dialogues involving multiple speakers fail to utilize information about who is speaking, leading to imprecise tagging and inability to set tags for utterance types, emotions, and paralinguistic information.
Innovation Solution
A tag estimation device that generates an utterance sequence information vector by combining utterance word and speaker vectors, using recurrent neural networks to determine tags for scenes, utterance types, and emotions, with model parameters learned from labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional scene tagging techniques are used for dialogues involving multiple speakers, then tagging can be performed on voice data, but speaker information cannot be utilized leading to imprecise tagging
Solution Approach 1:
The patent segments the dialogue processing by introducing speaker-specific feature extraction and speaker role identification as separate processing stages. Each speaker's utterances are analyzed individually with dedicated feature vectors, allowing the system to maintain distinct speaker information throughout the tagging process rather than treating all voice data uniformly.
Solution Approach 2:
The patent adds a new dimension to the tagging system by incorporating speaker identity and speaker role as additional feature dimensions. This transforms the tagging approach from analyzing only utterance content to analyzing a multi-dimensional space that includes speaker characteristics, enabling more precise tagging through enhanced contextual information.
2Measurement precision
If speaker information is incorporated into the tagging system, then tagging accuracy improves, but system complexity increases
Solution Approach 1:
The patent implements a universal recurrent neural network framework that handles multiple functions: utterance recognition, speaker identification, speaker role determination, and scene tagging. This multi-functional approach avoids the need for separate specialized systems for each task, managing complexity through a unified architecture that processes all information types through a single cohesive model.
Solution Approach 2:
The patent merges speaker feature extraction, speaker role classification, and scene tagging into an integrated processing pipeline. By combining these functions into a unified system that shares common components and processes information flow continuously, the patent reduces overall system complexity compared to having separate independent systems for each function.
3Adaptability or versatility
If multiple types of tags are set for dialogues, then analysis capability improves, but conventional techniques can only set scene identification tags
Solution Approach 1:
The patent creates a universal tagging system that can simultaneously generate multiple types of tags including scene identification tags, speaker role tags, and utterance type tags. The recurrent neural network framework is designed to handle diverse tagging objectives through a single unified architecture, enabling the system to adapt to different analysis requirements without requiring separate specialized systems.
Solution Approach 2:
The patent implements a dynamic tagging system where the types and numbers of tags can be flexibly adjusted based on the specific dialogue analysis needs. The system can adaptively determine which tags to generate and how many tags are appropriate for each situation, providing versatility while managing complexity through dynamic configuration rather than fixed rigid structures.
Data Source
AI summary
A tag estimation device capable of estimating, for an utterance made among several persons, a tag representing a result of analyzing the utterance is provided. The tag estimation device includes an utterance sequence information vector generation unit that adds a t-th utterance word feature vector and a t-th speaker vector to a (t-1)-th utterance sequence information vector ut-1 that includes an utterance word feature vector that precedes the t-th utterance word feature vector and a speaker vector that precedes the t-th speaker vector to generate a t-th utterance sequence information vector ut, where t is a natural number, and a tagging unit that determines a tag lt that represents a result of analyzing a t-th utterance from a model parameter set in advance and the t-th utterance sequence information vector ut.