Transcription
Bayanat can transcribe audio and video media to searchable text using Whisper, running locally on your own server. It is the audio/video counterpart to OCR: transcripts are stored alongside the media and become searchable.
How It Works
Transcription runs as part of Media Import. When you import audio or video files, you can opt to transcribe them, and a background worker generates the transcript automatically.
- Start a Media Import and add audio or video files
- Enable Transcribe audio/video files in the import options
- Optionally choose a language, or leave it blank for automatic detection
- Process the import; transcription runs asynchronously in the background
- The transcript is attached to each media item and marked as auto-generated
Models
Whisper offers several models, selected under System Administration → Whisper Model. Larger models are more accurate but slower and need more memory; smaller models are faster and lighter.
| Model | Relative size | Notes |
|---|---|---|
tiny, base | Smallest | Fast, modest accuracy. base is the default. |
small, medium | Mid | Better accuracy, slower. |
large | Largest | Best accuracy, slowest, highest memory use. |
English-only variants (.en, e.g. base.en) are available and can be more accurate for English-only material.
First run downloads the model
The selected model is downloaded and cached on the server the first time it is used. Later transcriptions reuse the cached copy. A GPU speeds transcription up considerably but is not required; Whisper also runs on CPU.
Languages
Whisper auto-detects the spoken language by default. You can also set a specific language for an import when you know it in advance, which can improve accuracy.
Reviewing and Correcting Transcripts
Auto-generated transcripts are a starting point, not a final record. Reviewers (Admin or Data Analyst roles) can correct the text directly on the media item. Edits are preserved with a history of changes, the original is not lost, and the corrected text becomes the searchable version.
Searching Transcripts
Transcribed text is indexed and searchable the same way as OCR text, so a phrase spoken in a video can be found through normal search.
Enabling Transcription
Transcription is off by default. An administrator enables it under System Administration (the "Allow Transcription of Media Files" setting) and selects a Whisper model. The transcription engine ships as an optional component; see Configuration for installation details.