How aws-samples/amazon-transcribe-live-meeting-assistant works
Project type and purpose
• This is a full-stack, serverless, event-driven AWS solution (“Live Meeting Assistant”) that captures live meeting audio (via a browser extension or direct stream), sends it to AWS services for transcription, translation, AI-driven queries and summarization, and presents real-time and post-meeting insights in a React web UI.
• Key use cases: live transcript with speaker attribution, live translation, contextual Q&A (via Amazon Q Business or Bedrock Knowledge Bases), on-demand and automated end-of-meeting summaries, meeting recordings, transcript history search.
File structure
• Top-level directories and modules:
• lma-ai-stack – main nested CloudFormation for core services (AppSync + GraphQL, Lambdas, Fargate task, Kinesis, DynamoDB, S3) and React UI under source/ui
• lma-websocket-transcriber-stack – WebSocket server in Fargate + supporting Lambdas
• lma-meetingassist-setup-stack – QnABot on AWS integration and custom prompts
• lma-virtual-participant-stack, lma-bedrockagent-stack, lma-bedrockkb-stack, etc. – optional feature stacks
• lma-browser-extension-stack – Chrome extension code
• Patterns: event-driven via Kinesis Data Streams, microservices style Lambdas, API layer via AppSync GraphQL, CI/CD via Makefiles and build scripts, Infrastructure as Code via CloudFormation templates.
README insights
• The “Processing flow overview” section outlines the end‐to‐end architecture: browser → WebSocket → Transcribe → Kinesis → Lambda “Call Event Processor” → AppSync & QnABot/Bedrock → React UI (Cognito auth, CloudFront).
• The README enumerates all AWS services, nested stacks, and data flows.
Main components & interactions
• Clients:
• Browser Extension (Chrome)
• React Web UI (hosted on S3 + CloudFront, authenticated via Cognito)
• Ingress:
• Secure WebSocket server (ECS Fargate behind ALB)
• GraphQL / AppSync API (WebSocket & HTTP)
• Streams & Storage:
• Amazon Transcribe (streaming)
• Amazon Kinesis Data Streams
• Amazon S3 (recordings, transcripts, build artifacts)
• DynamoDB (meetings metadata)
• Compute / Processing:
• Lambda: Call Event Processor
• Lambda: Transcript Summarization
• Lambda Hook functions (optional transcript enrichment)
• AI / Enrichment services:
• Amazon Translate
• Amazon Bedrock (LLMs & Knowledge Bases)
• Amazon Q Business / QnABot on AWS (OpenSearch)
• Amazon Lex (optional for orchestrator)
• Auxiliary services:
• Amazon Cognito (auth)
• CloudFront (static content)
• EventBridge (sync transcript KB)
• IAM roles & VPC
1. Browser Extension streams stereo audio + metadata → WebSocket server.
2. WebSocket task sends audio → Transcribe → transcription segments → Kinesis.
3. Kinesis → Call Event Processor Lambda: enrich, route to QnABot/Bedrock or custom hook.
4. Call Event Processor writes mutations to AppSync → DynamoDB & real-time updates to React UI subscriptions.
5. At call end: Summarization Lambda invokes Bedrock → summary stored in DynamoDB & S3; optionally populates transcript KB.
6. React UI queries AppSync for live segments, translations (via Translate), summaries, meeting list.
7. EventBridge periodically syncs S3 transcripts into Bedrock Knowledge Base for “Meetings Query Tool.”
c. Architectural patterns & design principles:
• Event-driven microservices: Kinesis streams, Lambdas for loose coupling and scaling.
• Serverless where possible (Lambdas, AppSync, Cognito), containerized WebSocket service for heavy audio processing.
• Nested CloudFormation stacks for modular, reusable infra.
• Least-privilege IAM roles per component.
• Extensibility via Lambda hook points and custom CloudFormation parameters.
d. Key technologies & frameworks:
• AWS CloudFormation, AWS SAM/Makefile
• AWS AppSync (GraphQL)
• AWS Lambda (Python & Node.js)
• ECS Fargate + ALB (Docker)
• React, AWS Amplify or aws-exports in UI
• Browser extension (TypeScript/JS)
• QnABot on AWS (OpenSearch + Bedrock agent)
• Transcribe, Translate, Bedrock, Q Business.