Speaker Diarization With Block Embeddings for Episodic Audio

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speaker diarization technologies struggle with efficiently partitioning audio streams containing speech from multiple individuals without requiring speaker fingerprints or prior knowledge of the number of speakers, and they are inadequate for handling overlapping speech.

Innovation Solution

A method involving spatial conversion, block-based embeddings extraction, multi-head attention architecture, embeddings optimization, and double clustering to enhance speaker diarization, allowing for accurate separation of speakers in diverse audio environments.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If speaker diarization is performed on entire audio streams without blocking, then speaker identification accuracy is maintained, but memory consumption and processing load become unmanageable

Engineering Contradiction:
Improvespeaker identification accuracyVSAvoidmemory consumption
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The audio stream is divided into fixed-size blocks that can be processed independently. Each block contains a segment of the audio stream that is diarized separately, allowing memory usage to be bounded while maintaining accuracy through proper block design and overlapping regions

Inventive Principle:
Principle #1Segmentation

2Stability of the object's composition

If speaker diarization is performed on entire audio streams without blocking, then continuous speaker tracking is improved, but processing time and computational load increase significantly

Engineering Contradiction:
Improvecontinuous speaker trackingVSAvoidprocessing time
Core Design Contradiction:
Stability of the object's compositionVSLoss of time

Solution Approach 1:

Processing is segmented into fixed-size blocks that can be handled independently and in parallel. This reduces processing time by allowing efficient block-wise computation while maintaining continuous speaker tracking through block chaining and overlap handling

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Blocks are pre-processed and prepared in advance, with speaker embeddings extracted and stored for later clustering. This preliminary action allows the main diarization logic to run more efficiently on pre-prepared data

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If embeddings are extracted for all segments before clustering, then comprehensive speaker analysis is achieved, but memory requirements become prohibitive

Engineering Contradiction:
Improvespeaker analysis comprehensivenessVSAvoidmemory requirements
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The audio stream is divided into fixed-size blocks that can be processed independently. Each block contains a segment of the audio stream that is diarized separately, allowing memory usage to be bounded while maintaining accuracy through proper block design and overlapping regions

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Blocks are pre-processed and prepared in advance, with speaker embeddings extracted and stored for later clustering. This preliminary action allows the main diarization logic to run more efficiently on pre-prepared data

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4330965B1Speaker diarization supporting eposodical content
Publication Date: 2026.04.22 DOLBY LABORATORIES LICENSING CORP
  • EP4330965B1 patent drawingFigure 1
  • EP4330965B1 patent drawingFigure 2A
  • EP4330965B1 patent drawingFigure 2B

AI summary

Embodiments are disclosed for speaker diarization supporting episodical content. In an embodiment, a method comprises: receiving media data including one or more utterances; dividing the media data into a plurality of blocks; identifying segments of each block of the plurality of blocks associated with a single speaker; extracting embeddings for the identified segments in accordance with a machine learning model, wherein extracting embeddings for identified segments further comprises statistically combining extracted embeddings for identified segments that correspond to a respective continuous utterance associated with a single speaker; clustering the embeddings for the identified segments into clusters; and assigning a speaker label to each of the embeddings for the identified segments in accordance with a result of the clustering. In some embodiments, a voiceprint is used to identify a speaker and the speaker identity for a speaker label.