Automatic Speaker Recognition in Meeting Audio Transcripts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automatic speech recognition systems fail to detect and recognize individual speakers in meetings, limiting the ability to match transcripts and recordings with speakers, search, filter, and derive insights from their contributions.
Innovation Solution
A machine learning model is applied to recordings to generate embeddings and labels for speakers based on metadata such as timestamps and voice activity, enabling automatic speaker recognition without manual user input, using a system that includes a processing engine, labeling engine, and analysis engine to identify and annotate speakers in meeting transcripts and recordings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automatic speech recognition engines are used to generate transcripts from meetings, then transcript generation is enabled, but the ability to detect and recognize individual speakers is lost
Solution Approach 1:
The audio recording is divided into multiple audio clips, each associated with a speaker label. This segmentation allows the system to process and identify individual speakers within the overall transcript generation process, resolving the contradiction between automated transcript generation and speaker identification.
Solution Approach 2:
A machine learning model acts as an intermediary between the audio recording and the transcript generation process. This model analyzes audio clips to generate speaker labels, which are then integrated into the transcript system, enabling both automated transcription and speaker recognition.
2Measurement precision
If manual user review and input are used to label audio with speakers, then accurate speaker recognition is achieved, but interactivity and functionality are reduced
Solution Approach 1:
The system performs automatic speaker recognition through machine learning models that analyze audio clips and generate speaker labels without requiring manual user review. This self-service approach maintains high speaker recognition accuracy while significantly improving system interactivity and ease of operation.
Solution Approach 2:
The system uses voice activity detection and audio feature analysis to automatically determine speaker characteristics. By changing the parameters from manual labeling to automated acoustic feature extraction, the system achieves both accurate speaker recognition and improved ease of operation.
3Difficulty of detecting and measuring
If speaker recognition functionality is added to meeting solutions, then speaker identification capability is improved, but device complexity increases
Solution Approach 1:
The speaker recognition system is segmented into distinct functional modules: audio clip extraction, voice activity detection, machine learning model processing, and label generation. This modular segmentation improves speaker detection capability while managing system complexity through organized, independent components.
Solution Approach 2:
The machine learning model serves multiple functions: it processes audio clips to generate speaker labels, identifies unique speakers in recordings, and integrates with the existing transcript generation system. This multi-functionality improves speaker detection capability without proportionally increasing system complexity.
Data Source
AI summary
One embodiment of the present invention sets forth a technique for processing recordings of events. The technique includes applying a machine learning model to a plurality of samples from one or more recordings of the events to generate a plurality of embeddings representing the plurality of samples. The technique also includes generating a plurality of labels that identify speakers associated with the plurality of samples based on metadata comprising timestamps of voice activity during the events and participants associated with the voice activity. The technique further includes storing mappings of the plurality of embeddings to the plurality of labels.