Skip to main content

What are Custom Models?

Custom language models allow you to train specialized speech recognition models tailored to your specific domain, improving transcription accuracy for your unique content.

Overview​

A custom model is a specialized version of Scriptix's base language models that has been trained with your own audio data. This training helps the model better recognize:

  • Domain-specific terminology - Medical, legal, technical, or industry-specific terms
  • Product and company names - Your organization's specific vocabulary
  • Accents and speaking styles - Particular speakers or regional accents
  • Audio environments - Specific recording conditions or setups

How Custom Models Work​

Base Models vs. Custom Models​

Base Models:

  • Pre-trained on general speech data
  • Work well for standard conversational speech
  • Available immediately for all users
  • Cover common vocabulary and patterns

Custom Models:

  • Built on top of a base language model
  • Trained with your specific audio and transcripts
  • Optimized for your domain and use case
  • Require training time and data preparation

The Training Process​

  1. Select Base Language - Choose which language to build upon (English, Dutch, French, etc.)
  2. Upload Training Data - Provide audio files with corresponding transcripts
  3. Train the Model - System learns from your data (takes several hours)
  4. Use in Production - Select your custom model when transcribing new audio

Benefits of Custom Models​

Improved Accuracy​

  • 15-30% improvement for specialized terminology
  • Fewer errors on domain-specific words
  • Better recognition of uncommon names and terms
  • Reduced post-transcription editing time

Consistency​

  • Same terminology spelled consistently
  • Product names recognized correctly every time
  • Standard formatting for specialized terms

Time Savings​

  • Less manual correction needed
  • Faster transcript finalization
  • More productive workflow

Domain Expertise​

  • Models understand your industry's language
  • Better handling of technical jargon
  • Improved recognition of field-specific patterns

Use Cases​

Medical & Healthcare​

What it helps with:

  • Medical terminology (diagnoses, procedures, medications)
  • Anatomical terms
  • Clinical abbreviations
  • Healthcare provider names

Example:

  • Before: "The patient has a my o cardial infarction"
  • After: "The patient has a myocardial infarction"

What it helps with:

  • Legal terminology and procedures
  • Case names and citations
  • Party names and titles
  • Court-specific vocabulary

Example:

  • Before: "The plaintiff filed a motion for summary judgment"
  • After: "The plaintiff filed a Motion for Summary Judgment"

Corporate & Business​

What it helps with:

  • Company and product names
  • Internal terminology and acronyms
  • Department and project names
  • Industry-specific jargon

Example:

  • Before: "We're using post gress sequel database"
  • After: "We're using PostgreSQL database"

Technical & IT​

What it helps with:

  • Software and technology names
  • Programming terminology
  • Technical specifications
  • System and tool names

Requirements​

Training Data​

To create a custom model, you need:

Audio Files:

  • High-quality recordings
  • Representative of your target content
  • Ideally 2-20 hours of audio
  • WAV, MP3, or other standard formats

Transcripts:

  • Accurate text transcriptions of the audio
  • Properly formatted and edited
  • Matching the audio exactly
  • Plain text format

Subscription Plan​

Custom models may require:

  • Specific subscription tier
  • Training capacity allocation
  • Storage for model data

Contact your organization administrator for details about custom model availability on your plan.

Comparison with Glossaries​

Custom models and glossaries both improve accuracy but work differently:

Custom Models​

Best for:

  • Large volumes of specialized content
  • Comprehensive domain coverage
  • Ongoing, regular use
  • Multiple types of terminology

Requires:

  • Training data (audio + transcripts)
  • Training time (hours to days)
  • Higher-tier plan

Glossaries​

Best for:

  • Specific terms and names
  • Quick implementation
  • Limited vocabulary (less than 500 terms)
  • Occasional use

Requires:

  • List of terms only
  • Immediate availability
  • Available on most plans

Recommendation: Use both together for maximum accuracy. Train a custom model for your domain, and add a glossary for specific product names or recent terminology.

Getting Started​

Ready to create a custom model?

  1. Assess Your Needs - Review When to Use custom models
  2. Prepare Data - Gather audio and transcripts
  3. Create Model - Follow the Create Custom Model guide
  4. Train - Learn about the Training Process
  5. Use - Start Using Your Model in transcriptions

Limitations​

What Custom Models Cannot Do:

  • Transcribe languages they weren't trained for
  • Handle completely new vocabulary without retraining
  • Work across different languages (each language needs its own model)
  • Improve audio quality issues

Training Limitations:

  • Requires minimum amount of quality training data
  • Training takes time (several hours to complete)
  • May need periodic retraining as vocabulary evolves

Next Steps​

Transform your transcription accuracy! Custom models bring domain expertise to speech recognition.