AI Summarization

OpenTranscribe generates AI-powered summaries of your transcripts using Large Language Models (LLMs).

Overview

AI summarization provides:

• BLUF Format: Bottom Line Up Front executive summaries
• Speaker Analysis: Talk time, key contributions by speaker
• Action Items: Extracted with priorities and assignments
• Key Decisions: Important conclusions and agreements
• Follow-up Items: Next steps and pending tasks

Requirements

• LLM provider configured (see LLM Integration)
• Completed transcription with or without speakers
• Sufficient LLM API credits (for cloud providers)

Generating Summaries

From Web UI

1. Open transcription details
2. Click "Generate Summary"
3. Select summary type (if custom prompts configured)
4. Wait for AI processing
5. View results in Summary tab

Automatic Processing

For long transcripts, OpenTranscribe automatically:

• Splits content into sections at speaker/topic boundaries
• Processes each section independently
• Stitches results into cohesive summary
• Handles transcripts of any length

Summary Formats

BLUF (Default)

Bottom Line Up Front format includes:

Executive Summary:

• 2-3 sentence overview
• Key takeaways
• Critical information first

Speaker Analysis:

• Talk time percentage
• Key contributions
• Speaking patterns

Action Items:

• Task description
• Priority (High/Medium/Low)
• Assigned person (if mentioned)
• Due date (if mentioned)

Key Decisions:

• Important conclusions
• Agreements reached
• Changes approved

Follow-up Items:

• Pending questions
• Future discussions
• Next steps

Custom Prompts

Create custom summarization prompts for specific use cases:

• Meeting notes format
• Interview analysis
• Podcast highlights
• Legal deposition summaries
• Medical consultation notes

LLM Providers

Local LLM (Privacy-First)

Best for sensitive content:

# vLLM or Ollama
LLM_PROVIDER=vllm
VLLM_API_URL=http://localhost:8000/v1

Advantages:

• Complete privacy
• No API costs
• Works offline
• Unlimited usage

Requirements:

• Dedicated GPU (8GB+ VRAM)
• Model deployment (Llama, Mistral, etc.)

Cloud LLM

Best for convenience:

# OpenAI
LLM_PROVIDER=openai
OPENAI_API_KEY=sk-xxxxx

# Anthropic Claude
LLM_PROVIDER=anthropic
ANTHROPIC_API_KEY=sk-ant-xxxxx

Advantages:

• No infrastructure needed
• Latest models
• High quality results

Considerations:

• Requires internet
• Per-token costs
• Data sent to provider

Performance

Transcript Length	Processing Time (Cloud LLM)	Processing Time (Local LLM)
30 minutes	10-15 seconds	20-40 seconds
1 hour	20-30 seconds	40-80 seconds
3 hours	60-90 seconds	2-4 minutes

Note: OpenTranscribe intelligently chunks long transcripts, so a 10-hour transcript takes only marginally longer than a 3-hour one.

Configuration

Provider Settings

Edit .env:

# Provider selection
LLM_PROVIDER=vllm  # or: openai, anthropic, ollama, openrouter

# Provider-specific
VLLM_API_URL=http://localhost:8000/v1
VLLM_MODEL_NAME=meta-llama/Llama-2-70b-chat-hf
OPENAI_API_KEY=sk-xxxxx
ANTHROPIC_API_KEY=sk-ant-xxxxx

Custom Prompts

Create custom prompts in UI:

1. Go to Settings → AI Prompts
2. Click "New Prompt"
3. Configure:
   • Name
   • System instructions
   • JSON schema (for structured output)
   • Temperature
4. Save

Use custom prompts:

1. Generate Summary
2. Select your custom prompt
3. Process

Best Practices

1. Review AI Summaries: Always verify critical information
2. Speaker Labels Help: Labeled speakers produce better summaries
3. Clear Audio: Better transcription = better summaries
4. Choose Right Provider: Local for privacy, Cloud for quality
5. Custom Prompts: Tailor summaries to your workflow

Troubleshooting

Summary Generation Fails

Check:

• LLM provider configured correctly
• API key valid (for cloud)
• LLM server running (for local)
• Sufficient credits (for cloud)

Solution:

# Test LLM connection
./opentr.sh logs celery-worker | grep -i llm

# Verify provider settings
grep LLM_ .env

Poor Quality Summaries

Causes:

• Weak LLM model
• Poor transcription quality
• Insufficient context

Solutions:

• Use larger model (70B+ parameters recommended)
• Improve transcription (use large-v2 Whisper model)
• Add speaker labels for better context

Slow Processing

Solutions:

• Use cloud LLM (faster)
• Upgrade local LLM hardware
• Use smaller model (trade-off with quality)

Cost Estimates (Cloud)

Approximate costs for cloud LLM providers:

Transcript Length	OpenAI GPT-4	Claude Opus	OpenRouter
30 minutes	$0.20-0.40	$0.25-0.50	$0.15-0.30
1 hour	$0.40-0.80	$0.50-1.00	$0.30-0.60
3 hours	$1.20-2.40	$1.50-3.00	$0.90-1.80

Note: Local LLM has zero per-use cost after hardware investment.

Next Steps

• LLM Integration
• Speaker Management
• Search & Filters