Skip to content

Transcription Service

The Transcription tool uses WhisperX to convert spoken audio into written text with precise timestamps and speaker identification.

Supported Formats

You can upload audio files in the following formats:

  • .wav
  • .mp3
  • .flac
  • .m4a
  • .zip (for batch processing multiple audio files at once)

How to Use

Single Audio Transcription

  1. Select Workflow: Choose "Transcribe audio".
  2. Upload File: Drag and drop your audio file.
  3. Language Selection:
  4. The system will attempt to auto-detect the language.
  5. Important: You must verify or manually select the correct language from the dropdown.

Batch Processing (Multiple Files)

If you have multiple audio recordings (e.g., a folder of 10 interviews), you can transcribe them all in one go without clicking through them individually:

  1. Select Workflow: Choose "Batch transcribe (ZIP)".
  2. Upload File: Compress all your audio files into a single .zip archive on your computer and upload it.
  3. Language Selection: * You can choose a specific language to force the AI to use that language for all files.
  4. Alternatively, select Auto-detect. The AI will dynamically analyze and figure out the correct language for each file individually before transcribing it.
  5. Process: Click Start. The AI will load the models once and loop through your entire ZIP file at maximum speed.

🇨🇭 Swiss German Support

Text Lab includes a specialized fine-tuned model for Swiss German.

  • Select Swiss German from the language list.
  • This uses the swhisper-large-1.1 model, which achieves a Word Error Rate (WER) of approximately 18, significantly outperforming standard models on Swiss German dialects. More info about the fine-tuned whisper will be published soon.

Configuration Options

  • VAD (Voice Activity Detection) (Optional):
  • Check this to filter out silence before processing.
  • Max Pause: Determines how much silence is allowed between words before splitting a segment.
  • Speaker Diarization (Who is speaking?):
  • Min/Max Speakers (Optional): If you know how many people are in the recording, set these numbers to help the AI distinguish between voices (e.g., for an interview, set Min=2, Max=2).

Results & Export

Once the transcription is complete, you can:

  • Preview (Single Files Only): Listen to the audio with a synchronized interactive text player.
  • Edit View: Toggle between "Segments" (sentences) and "Words" view.
  • Download:
  • Text (.txt): Plain text transcript.
  • CSV: Contains timestamps and speaker labels.
  • Subtitles (.srt & .vtt): Ready-to-use subtitle files that can be imported directly into video editing software (Premiere Pro, DaVinci Resolve), VLC media player, or YouTube.
  • ZIP Package: Downloads all the formats above in one convenient package. (For batch processing, this ZIP will automatically organize your transcripts into individual, neatly named folders for each audio file!)

Data Privacy & Security

Audio recordings (such as interviews, meetings, or field notes) often contain highly sensitive personal data. Text Lab processes all audio using a strict "Zero-Footprint" architecture to ensure absolute confidentiality.

  • 100% Local Processing: Unlike commercial services (e.g., standard OpenAI Whisper APIs, Otter.ai, etc.), your audio files are never sent to the cloud. All transcription is performed locally on the University of Bern's secure UBELIX high-performance computing nodes.
  • Volatile Memory Storage: When you upload an audio file, it is loaded into the server's temporary volatile memory (RAM).
  • Secure Temporary Decoding: If the app needs to convert complex formats (like .m4a or .mp3), it may briefly write a hidden temporary file to the node's secure storage. This file is instantly and permanently deleted the moment the decoding is finished—usually within milliseconds.
  • Ephemeral Sessions: The generated transcripts and the audio waveforms loaded into the browser's playback tool are tied strictly to your active session. When you close the browser tab or your HPC job ends, all audio and text data are instantly wiped. Your data is never saved to your university home directory.
  • No AI Training: The Whisper models only perform inference (listening and transcribing). They do not learn from your voice recordings, and your data is never used to train or improve the AI.