Transcription Backends¶

Voice Transcriber supports two transcription backends: OpenAI Whisper (cloud) and Speaches (self-hosted).

Backend Comparison¶

OpenAI Whisper (Cloud)¶

Best for: Quick setup, occasional use, no local resources

Configuration¶

{
  "language": "en",
  "transcription": {
    "backend": "openai"
  }
}

Pros¶

• ✅ Zero setup required
• ✅ No local resources needed
• ✅ Consistently fast processing
• ✅ High accuracy across languages

Cons¶

• ❌ Requires internet connection
• ❌ API costs ($0.006 per minute of audio)
• ❌ Audio data sent to OpenAI servers
• ❌ Subject to OpenAI API rate limits

Speaches (Self-Hosted)¶

Best for: Privacy-conscious users, high-volume use, offline operation

Powered by Speaches - OpenAI-compatible speech-to-text server

Quick Setup (3 commands)¶

# 1. Create docker-compose.speaches.yml
docker compose -f docker-compose.speaches.yml up -d

# 2. Update config
nano ~/.config/voice-transcriber/config.json
# Change "backend": "openai" to "backend": "speaches"

# 3. Done! First transcription downloads model (~140MB)

Configuration¶

{
  "language": "fr",
  "formatterEnabled": false,
  "transcription": {
    "backend": "speaches",
    "speaches": {
      "url": "http://localhost:8000/v1",
      "apiKey": "none",
      "model": "Systran/faster-whisper-base"
    }
  }
}

Pros¶

• ✅ Zero cost - No API fees
• ✅ Complete privacy - Audio never leaves your machine
• ✅ Offline operation - No internet required after model download
• ✅ Same speed - Base model comparable to OpenAI (3.7s vs 3.8s)
• ✅ High accuracy - 91-100% similarity depending on model

Cons¶

• ❌ Requires Docker setup
• ❌ Initial model download (~140MB-2.9GB depending on model)
• ❌ Requires local compute resources
• ❌ Larger models need more RAM/CPU

Performance Comparison¶

Benchmark: 30s French audio, Remote server (8 CPU / 8GB RAM)

Choosing a Backend¶

Use OpenAI Whisper if:¶

• 📱 You want zero setup and immediate use
• 🌐 You always have internet connection
• 💵 Cost is acceptable for your usage volume
• 🎯 You prioritize convenience over privacy

Use Speaches if:¶

• 🔒 Privacy is important (audio never leaves your machine)
• 💰 You transcribe frequently (avoid API costs)
• 📴 You need offline operation
• 🏠 You have local compute resources (or can spin up a VPS)

Switching Backends¶

OpenAI → Speaches¶

# 1. Setup Speaches with Docker
docker compose -f docker-compose.speaches.yml up -d

# 2. Update config
nano ~/.config/voice-transcriber/config.json

Change:

{
  "transcription": {
    "backend": "speaches"
  }
}

Restart the application.

Speaches → OpenAI¶

# Update config
nano ~/.config/voice-transcriber/config.json

Change:

{
  "language": "en",
  "formatterEnabled": true,
  "transcription": {
    "backend": "openai",
    "openai": {
      "apiKey": "sk-your-api-key-here"
    }
  }
}

Restart the application.

Benchmark Mode¶

Compare both backends side-by-side. Requires both OpenAI and Speaches configured.

Configuration¶

{
  "benchmarkMode": true,
  "transcription": {
    "backend": "speaches",
    "openai": {
      "apiKey": "sk-...",
      "model": "whisper-1"
    },
    "speaches": {
      "url": "http://localhost:8000/v1",
      "apiKey": "none",
      "model": "Systran/faster-whisper-base"
    }
  }
}

Running Benchmarks¶

Run with --debug flag to see detailed comparison:

voice-transcriber --debug

Benchmark Output¶

Example output:

🔬 BENCHMARK: Comparing OpenAI and Speaches
⏱️  Performance:
   OpenAI Whisper:   2.45s
   Speaches:         0.87s
   Speedup:          2.82x faster

📏 Text Length:
   OpenAI:   142 chars
   Speaches: 145 chars
   Difference: 3 chars (2.1%)

🎯 Similarity: 97.2% match

What Benchmark Mode Does¶

When enabled, the application:

1. Transcribes with both backends simultaneously
2. Measures performance - Processing time for each backend
3. Compares accuracy - Text similarity calculation between results
4. Shows differences - Character count and text length comparison
5. Uses primary backend result - The configured backend result is copied to clipboard

Use Cases¶

Disabling Benchmark Mode¶

Set to false in config:

{
  "benchmarkMode": false
}

Restart the application for changes to take effect.

Next Steps¶

• Speaches Integration Guide - Detailed setup
• Whisper Models Comparison - Model selection
• Configuration Guide - Advanced settings