Skip to main content

Custom Models

Custom models are specialized speech recognition models that can be trained to improve transcription accuracy for specific vocabulary and terminology.

Overview​

Custom models optimize speech recognition for specific terminology and industry vocabulary. Available on Business (Gold, €1,199/month) and Enterprise (Platinum, €10,000/month) plans.

Important: Custom models are organization-specific. No verified cross-organization sharing functionality exists.

Requirements: 1-10 dataset files (10GB max per file). Supported formats: Audio (.wav, .mp3, .m4a, .flac), Transcript (.vtt, .srt, .txt), Manifest (.jsonl)

Creating Custom Models​

Navigate to Custom Models → "Create Custom Model" → fill form (Name*, Language*) → Create. Language selection determines base model. Guide modal appears on first visit.

Training Custom Models​

Click model name → "Add Dataset" → select type (Auto/Audio/Transcript/Manifest) → upload files (1-10, 10GB max) → submit. Dataset table shows: Name, Type, Duration, URL, Start/End time.

Train Model: Click "Train" → confirm → training begins.

Training Status (5 states): Not Running (Zinc), Ready to Run (Amber), Running (Sky), Success (Emerald), Failed (Red). Train button available when Not Running or Failed.

Managing Custom Models​

Models List: View table with ID, Name, Base Language, Training Status, Last Modified, Type. Filter by language/status, sort, search by name.

Model Details: Click model name → view/edit name, language, last modified, organization ID. Save changes.

Delete: ORGADMIN and SYSOP only. Click Delete → confirm → deleted.

Permissions: ORGADMIN (create, train, update, delete, view all), SYSOP (full access, back-office), Other Roles (view only, cannot delete).

Organization Ownership: Each model belongs to one organization. No cross-organization sharing.

Using Custom Models​

Select trained custom model during transcription upload/setup for improved accuracy on domain-specific content.

Best Practices​

Training Data: Use clear, high-quality audio with minimal background noise. Provide accurate word-for-word transcripts matching audio files. Upload 1-10 dataset files (10GB max each).

Model Management: Use descriptive names (e.g., "Medical_EN_US"). Wait for training completion before use. Test with sample transcriptions and compare accuracy.

Troubleshooting​

Training Fails: Check file formats/sizes, ensure audio-transcript match, review dataset quality, retry training.

Cannot Delete: Requires ORGADMIN or SYSOP role.

Upload Fails: File size limit 10GB. Check valid formats (Audio: .wav/.mp3/.m4a/.flac, Transcript: .vtt/.srt/.txt, Manifest: .jsonl).

Next Steps​