Skip to main content

Training Process

Learn how to train your custom language model with your audio and transcript data.

Training Overview​

Training is the process of teaching your custom model to recognize speech patterns, vocabulary, and characteristics specific to your domain. This requires:

  1. Training Data - Audio files with matching transcripts
  2. Data Upload - Adding your files to the model
  3. Training Execution - Running the training process
  4. Validation - Confirming successful training

Before Training​

Data Requirements​

Audio Files:

  • Format: WAV, MP3, M4A, or other standard audio formats
  • Quality: Clear recordings with minimal background noise
  • Duration: Ideally 2-20 hours total
  • Content: Representative of your target transcriptions

Transcript Files:

  • Format: Plain text (.txt)
  • Content: Exact transcription of the audio
  • Accuracy: Must match audio precisely
  • Encoding: UTF-8 recommended

Audio-Transcript Pairing:

  • Each audio file must have a corresponding transcript
  • Filenames should match (e.g., recording1.mp3 + recording1.txt)
  • Content must align exactly

Data Preparation Tips​

Audio Quality:

  • Use high-quality recordings when possible
  • Consistent recording environment
  • Clear speech, minimal overlapping
  • Similar audio characteristics to your target use case

Transcript Accuracy:

  • Transcripts must be highly accurate (95%+ accuracy)
  • Include all spoken words
  • Proper punctuation and capitalization
  • Match the audio exactly, word for word

Data Quantity:

  • More data generally = better results
  • Minimum: Usually 1-2 hours of audio
  • Recommended: 5-10 hours for good results
  • Optimal: 10-20+ hours for best results

Accessing Training Interface​

  1. Navigate to Custom Models page
  2. Click on your model name or Edit icon (pencil)
  3. Model detail page opens
  4. Training interface displays

Training Status States​

Your model progresses through these states:

1. Not Running (Gray Badge)​

What it means:

  • Model created but no training started
  • No training data uploaded yet

What to do:

  • Upload training data (audio + transcripts)
  • Prepare to start training

2. Ready to Run (Amber Badge)​

What it means:

  • Training data has been uploaded
  • Model is ready to begin training
  • Waiting for you to start the process

What to do:

  • Review uploaded data
  • Click "Start Training" or similar button
  • Training will begin

3. Running (Blue Badge)​

What it means:

  • Training is actively in progress
  • System is processing your data
  • Model is learning from your audio and transcripts

What to do:

  • Wait for training to complete
  • Training typically takes several hours
  • Monitor progress if status updates are available
  • Do not interrupt the process

Duration:

  • Depends on data volume
  • Typically 2-12 hours
  • Larger datasets take longer

4. Success (Green Badge)​

What it means:

  • Training completed successfully
  • Model is ready to use
  • Can be selected for transcription

What to do:

  • Model is now ready for production use
  • Start using it for transcriptions
  • Test accuracy with sample audio

5. Failed (Red Badge)​

What it means:

  • Training encountered an error
  • Model is not ready to use
  • Issue needs investigation

What to do:

  • Check error messages or logs (if available)
  • Review training data quality
  • Contact support if needed
  • May need to retry training

Uploading Training Data​

Via Model Detail Page​

  1. Open your model (click Edit or model name)
  2. Look for training data upload section
  3. Upload your files:
    • Audio files
    • Corresponding transcript files
  4. System validates and processes files
  5. Status changes to "Ready to Run" when upload complete

Data Organization​

Best Practice:

  • Organize files in pairs
  • Use clear, matching filenames
  • Example:
    • session01.mp3 + session01.txt
    • meeting_02.wav + meeting_02.txt
    • call_003.mp3 + call_003.txt

Starting Training​

Once data is uploaded and status is "Ready to Run":

  1. Review uploaded data
  2. Click "Start Training" or "Train Model" button
  3. Confirm training start (if prompted)
  4. Status changes to "Running" (blue)
  5. Training begins processing

Note: You cannot stop training once started. Ensure everything is ready before starting.

During Training​

What Happens​

The system:

  1. Processes your audio files
  2. Analyzes the matching transcripts
  3. Learns patterns and vocabulary
  4. Optimizes the model for your domain
  5. Validates the training results

Monitoring Progress​

Status Updates:

  • Check the Training Status badge
  • May show progress percentage (if available)
  • Estimated completion time (if shown)

What You Can Do:

  • Monitor status periodically
  • Continue other work
  • Do not delete the model while training
  • Wait for completion

Training Duration​

Typical Timelines:

  • Small dataset (1-3 hours audio): 2-4 hours training
  • Medium dataset (5-10 hours audio): 4-8 hours training
  • Large dataset (10-20 hours audio): 8-12+ hours training

Factors Affecting Duration:

  • Amount of training data
  • Audio file sizes
  • System load
  • Model complexity

After Training​

Successful Training​

When status shows "Success" (green):

  1. Model is ready to use
  2. Available for selection in transcription
  3. Will improve accuracy for your domain

Next Steps:

  • Test the model with sample audio
  • Compare accuracy to base model
  • Start using in production transcriptions

Failed Training​

If status shows "Failed" (red):

  1. Check for error messages in model details
  2. Review training data quality
  3. Common issues:
    • Audio-transcript mismatch
    • Poor audio quality
    • Insufficient data
    • File format issues

Troubleshooting:

  • Verify audio and transcript pairs match
  • Check file formats are supported
  • Ensure transcripts are accurate
  • Try with different/better data
  • Contact support with error details

Retraining​

When to Retrain​

Consider retraining your model when:

  • Vocabulary has evolved significantly
  • New terminology emerged
  • Accuracy has decreased
  • You have additional quality training data
  • Quarterly or bi-annual maintenance

How to Retrain​

  1. Gather new/additional training data
  2. Open your existing model
  3. Upload new data
  4. Start training again
  5. New training replaces previous version

Note: Retraining overwrites the previous model. Test before deploying to production.

Best Practices​

Data Quality​

Ensure High Quality:

  • Accurate transcripts (95%+ accuracy)
  • Clear audio recordings
  • Consistent audio quality
  • Representative of target use case

Incremental Approach​

Start Small, Expand:

  1. Begin with 2-5 hours of best quality data
  2. Train and test model
  3. Evaluate accuracy improvements
  4. Add more data if needed
  5. Retrain with expanded dataset

Testing​

Validate Training Success:

  1. Use model on test audio
  2. Compare to base model results
  3. Measure accuracy improvement
  4. Verify domain-specific terms recognized
  5. Gather user feedback

Maintenance​

Regular Updates:

  • Review model performance quarterly
  • Add new terminology as needed
  • Retrain with fresh data periodically
  • Keep training data organized and backed up

Troubleshooting​

Training Won't Start​

Problem: Can't start training despite uploading data

Solutions:

  • Verify status is "Ready to Run"
  • Check all data uploaded correctly
  • Ensure audio-transcript pairs complete
  • Try refreshing the page

Training Stuck at "Running"​

Problem: Training status stuck, not progressing

Solutions:

  • Wait longer (training takes hours)
  • Refresh page to check for updates
  • Contact support if stuck for 24+ hours

Training Failed​

Problem: Status shows "Failed"

Solutions:

  • Review error messages
  • Check training data quality
  • Verify file formats supported
  • Ensure audio-transcript matching
  • Try with different data
  • Contact support with details

Model Not Available After Success​

Problem: Training succeeded but model not usable

Solutions:

  • Refresh the page
  • Check model list
  • Verify status is green "Success"
  • Contact support if issue persists

Next Steps​

After successful training:

  • Use in Transcription - Apply your trained model
  • Test with sample audio
  • Compare accuracy improvements
  • Deploy to production use

Training complete! Your custom model is ready to improve transcription accuracy.