Skip to main content

Training Process

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

Training Overview​

Training teaches your custom model to recognize speech patterns and vocabulary specific to your domain. The process requires:

  1. Training Data - Audio and/or transcript files
  2. Data Upload - Adding files to the model
  3. Validation - System validates datasets
  4. Training Execution - Running the training process

Training Status States​

Models progress through five training states:

1. Not Running (Zinc Badge)​

What it means:

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

What to do:

  • Upload datasets (audio and/or transcripts)
  • Status will change when datasets uploaded

2. Ready to Run (Amber Badge)​

What it means:

  • Datasets have been uploaded
  • Model ready to begin training
  • Training queued to start

What appears:

  • "Train the Language" button not shown
  • Training will begin automatically when ready

3. Running (Sky Badge)​

What it means:

  • Training actively in progress
  • System processing your data
  • Model learning from datasets

What to do:

  • Wait for training to complete
  • Do not delete model during training
  • Monitor status periodically

What appears:

  • "Train the Language" button hidden
  • Cannot modify model during training

4. Success (Emerald Badge)​

What it means:

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

What to do:

  • Use model in STT Session
  • Select from language dropdown when creating transcripts

What appears:

  • "Train the Language" button hidden
  • Model available for use

5. Failed (Red Badge)​

What it means:

  • Training encountered an error
  • Model not ready to use

What to do:

  • Upload new datasets
  • Try training again
  • Check error messages

What appears:

  • "Train the Language" button visible
  • Can upload new datasets and retry

Uploading Datasets​

Access Model Details​

  1. Navigate to Custom Models page
  2. Click model name or Edit action
  3. Model details page opens
  4. Datasets section displays at bottom

Upload New Dataset​

  1. Click "New Dataset" button
  2. Modal opens
  3. Select dataset type:
    • Auto-detect (default)
    • TRANSCRIPT (.vtt, .srt, .txt)
    • AUDIO (.wav, .mp3, .m4a, .flac)
  4. Upload files:
    • Click upload area or drag-and-drop
    • Select 1-10 files
    • Maximum 10GB per file
  5. Click "Create"
  6. Files upload and appear in datasets list

Dataset Types​

TRANSCRIPT (1):

  • Format: .vtt, .srt, .txt
  • Text transcripts with or without timestamps
  • Used for training text patterns

TEST (2):

  • Format: .vtt, .srt, .txt
  • Test transcripts for validation
  • Verifies training quality

AUDIO (3):

  • Format: .wav, .mp3, .m4a, .flac
  • Audio files for training
  • Used for acoustic model training

MANIFEST (4):

  • Format: .jsonl
  • Organizes datasets
  • Links audio and transcripts

Upload Validation​

Files validated on upload:

  • File count: 1-10 files per upload
  • File size: Maximum 10GB per file
  • Format validation: Must match selected type
  • Auto-detection: Determines type from extension if "auto" selected

Validation errors:

  • validation_failed - Datasets failed validation
  • invalid_audio_format - Audio files invalid format
  • invalid_transcript_format - Transcript files invalid format
  • File size exceeds 10GB limit

Starting Training​

Train Button Visibility​

"Train the Language" button appears only when:

  • Training status is Not Running (1) OR Failed (5)
  • Button hidden for Ready to Run (2), Running (3), Success (4)

Button Location:

  • Model details page header
  • Next to "Update" button
  • Only visible for applicable statuses

Training Process​

  1. Upload datasets if not already uploaded
  2. Click "Train the Language" button (if status is 1 or 5)
  3. Confirmation dialog appears:
    • Title: "Train the language confirmation"
    • Message: "Are you sure you want to start the training process for [model name]? This process may take a long time to complete and it will use your credits, and you will not be able to modify the custom model until it is finished."
  4. Click "Train" to confirm
  5. Validation runs - System validates datasets
  6. Training starts if validation passes
  7. Status changes to "Running" (3)
  8. Success toast: "Training started successfully"

Training Requirements​

Before training can start:

  • Datasets uploaded to model
  • Training credits available in organization
  • Model validation passes
  • Model not currently training
  • Model status allows training (1 or 5)

Error Codes:

  • validation_failed - Datasets validation failed
  • invalid_audio_format - Invalid audio file format
  • invalid_transcript_format - Invalid transcript format
  • no-claims-available - No training credits available

During Training​

What Happens​

The system:

  1. Validates all uploaded datasets
  2. Processes audio files
  3. Analyzes transcripts
  4. Trains the model
  5. Updates training status

Monitoring Progress​

Status Badge:

  • Shows "Running" (sky blue badge)
  • Located in model information cards

What You Can Do:

  • View model details
  • Monitor training status
  • Continue other work
  • Cannot modify model during training
  • Cannot delete model during training

What You Cannot Do:

  • Start training again (button hidden)
  • Upload new datasets (modal unavailable)
  • Modify model name during training
  • Delete model while training

Training Confirmation​

Confirmation modal warns:

  • Training may take a long time
  • Will use training credits
  • Cannot modify model until finished
  • Requires user confirmation before starting

After Training​

Successful Training (Status 4)​

When training completes successfully:

Model Details Page:

  • Status badge shows "Success" (emerald green)
  • "Train the Language" button hidden
  • Info banner shows success message with model name
  • Model ready for use

Status Help Message:

  • "Training complete! Select '[model name]' from the language dropdown when uploading files to use this model."

Next Steps:

  1. Navigate to workspace (Home page)
  2. Click "STT Session"
  3. Select language
  4. Select custom model from dropdown
  5. Upload audio
  6. Start transcription

Failed Training (Status 5)​

If training fails:

Model Details Page:

  • Status badge shows "Failed" (red)
  • "Train the Language" button appears
  • Can upload new datasets
  • Can retry training

Status Help Message:

  • "Upload datasets to get started. Add audio files and/or transcripts. Status will change to Ready to Run."

What to Do:

  1. Check error messages
  2. Review dataset quality
  3. Upload new/corrected datasets
  4. Click "Train the Language" to retry
  5. Confirm training start

Force Alignment for Training Data​

Use Force Alignment to prepare training data:

What It Does:

  • Adds timestamps to plain text transcripts
  • Creates properly formatted training files
  • Uses audio + existing transcript

How to Use:

  1. Navigate to workspace
  2. Click "STT Session"
  3. Select Force Alignment option
  4. Upload audio file
  5. Provide transcript text
  6. Process alignment
  7. Download result with timestamps
  8. Upload to custom model as training data

Dataset Management​

View Datasets​

Datasets List:

  • Displayed on model details page
  • Shows all uploaded datasets
  • Columns: Name, Type, Duration, URL, Start Time, End Time

Dataset Types Shown:

  • TRANSCRIPT
  • TEST
  • AUDIO
  • MANIFEST

Delete Datasets​

  1. Locate dataset in list
  2. Click three-dot menu on row
  3. Select "Delete"
  4. Confirmation dialog appears
  5. Confirm deletion
  6. Dataset removed from model

Dataset Deletion:

  • Available for all datasets
  • Permanent action
  • Cannot undo
  • Can delete during any training status

Training Credits​

Training requires credits:

Credits System:

  • Organization must have training credits
  • Credits consumed when training starts
  • Error if no credits available: "No training credits available"

Checking Credits:

  • Organization settings or billing page
  • Contact organization administrator
  • Error message displays if insufficient

Info Banner Messages​

Model details page shows status-specific help:

Status 1 (Not Running):

  • "Upload your training datasets (audio + transcripts), then click 'Train the language' to start."

Status 2 (Ready to Run):

  • "Training is queued and will begin shortly."

Status 3 (Running):

  • "Training in progress. This may take several hours depending on dataset size."

Status 4 (Success):

  • "Training complete! Select '[model name]' from the language dropdown when uploading files to use this model."

Status 5 (Failed):

  • Same as Status 1 message

Banner Features:

  • Located below model information cards
  • Expands by default when status is Success (4)
  • Collapsed by default for other statuses
  • Includes link to Force Alignment feature

Limitations​

Cannot Train When:

  • Status is Ready to Run (2)
  • Status is Running (3)
  • Status is Success (4)
  • No training credits available
  • Datasets fail validation

Training Process:

  • Cannot stop once started
  • Cannot modify model during training
  • Uses training credits
  • Requires validation to pass

Button Visibility:

  • Only visible for Not Running (1) or Failed (5)
  • Hidden for all other statuses

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

Next Steps​

After successful training:

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