4.3.3. Language Model Training Best Practices

• Concept: Training data quality affects transcription accuracy
• Purpose: Optimize for your domain vocabulary
• Benefit: Accurate transcription of specialized terms

When training a custom language model:

• ✅ Include multiple examples of spoken sentences
• ✅ Provide multiple adaptation options
• ✅ Put only one sentence per line
• ❌ Avoid repeating identical sentences (creates bias)
• ❌ Avoid 500,000+ sentences (dilutes boosting effect)
• ❌ Avoid special characters (~, #, @, %, &) - they're discarded