Transcriber User Guide

Approved for data classification level

0 1 2 3

Download Transcriber User Guide (PDF)

---

1. Using the Transcriber application

1.1 Finding the application

• Go into the application and use the search function to find Transcriber.

• Click on Transcriber.

1.2 Using the application

You should now see the following screen:

There are several options here, and it can seem overwhelming. For this example, we’ll walk through the quickest way to start a transcription.

1.2.1 Choose a name for your job

• Pick a name that makes it easy to find your data later and distinguish between different jobs.
• Example: "Transcriber demo 1".

Note: Job and file names cannot include special characters such as "æøå".

1.2.2 Select the duration of your job

• The application can transcribe in 1:1 time. For instance, a 1-hour audio file will take approximately 1 hour to transcribe.
• We recommend allocating double the length of the audio file to avoid interruptions.
• Example: For a 1-hour audio file, allocate 2 hours.

Note: If you run out of allocated time, the file being transcribed will fail. You can allocate more time after starting the job if needed.

1.2.3 Pick a machine to use

• We recommend the u3-gpu-1 machine, which performed best in our tests. If the option is unavailable we recommend the u1-standard-16 as an alternative.

Feel free to test with sample files to see what works best for you.

1.2.4 Select the input file

• Click the "use" button.

• Click the text box to select your file.

• Navigate to your "drives" and select the folder with your file or click "use" if it's already listed.

Note: The app can only process .mp3, .mp4, .m4a, .wav, and .mpg files. If your file is in another format, we recommend using VLC to convert it. VLC can be downloaded from the Software Center/Company Portal.

1.2.5 Select the output directory

• Choose where your output will be saved.
• Click "use" on "option: --output_dir".

• Select the folder you want for your transcription output.

Note: The app supports .mp3, .mp4, .m4a, .wav, and .mpg files. For other formats, consider converting using VLC.

Now, you are ready to begin your transcription. Click Submit to start the process.

There are additional options available. These are covered in the "Other Options" section.

Once the process starts, you can close your computer. If you want to ensure everything is running smoothly, wait until a "node" is assigned. Your screen will look like this:

Once the transcription is complete, you will see the following screen:

Note: This is not the actual output of your transcription. The transcription files are located in the folder you selected for output. You’ll find something like this:

---

1.3 Transcription output formats

You will have several different files with your transcription. Commonly used formats include .txt and .docx. You can choose the format that suits your needs best. If you want a specific output format, refer to the "output_format" section under "Other Options".

---

2. Optional parameters

2.1 Option: --output_format

By default, the application produces all 8 formats. You can limit the output to a specific format by selecting one of the following:

• CSV: Contains all parameters outputted from the Whisper model.
• SRT: SubRip file format, a widely adopted subtitle format.
• TXT: Pure text file with the transcription.
• VTT: Web Video Text Tracks format, includes timestamps.
• JSON: JavaScript Object Notation.
• TSV: Tab-separated values file containing start, end, and text.
• DOTE: Transcription software developed by the BigSoftVideo team at AAU.
• DOCX: Text file with transcription and speaker recognition.

2.2 Option: --output_model

Select the model size:

• Small: Faster but less accurate.
• Medium: Slightly slower, more accurate.
• Large: Most accurate but slowest.

The default is Large. With a machine featuring 16 vCPUs and 96GB of memory, transcription speed is about the same as the audio length (e.g., 1 minute of audio takes approximately 1 minute to transcribe).

2.3 Option: --output_language

Specify the language for transcription. The Whisper model can detect and automatically choose the language. If you select a language manually, the model will translate audio into that language.

Note: The detected or chosen language determines the output language. For example, if the chosen language is English, the model will translate multiple languages into English.

2.4 Interactive mode

Enable interactive mode for access to the app terminal or a web interface. The web interface includes a JupyterLab workspace for working with notebooks.

2.5 Archive password

Encrypt and password-protect the ZIP output archive. Specify a password for the archive as a text string.

2.6 Minimum and maximum number of speakers

Specify the number of speakers to improve speaker diarization accuracy in some cases.

2.7 Merge consecutive text entries from the same speaker (Recommended)

This option combines consecutive text entries from the same speaker into a single block, improving readability.

• When enabled, the app generates additional files with merged text in docx, dote, json, and csv formats. These files are named filename_merged and are created alongside the original files.

To make the option visible, scroll down in the optional parameter window.

---

For further assistance, contact CLAAUDIA at https://serviceportal.aau.dk.

---

Who made it?

Research & development by

• Nikolaj Andersen, CLAAUDIA, ITS, AAU
• Pelle Rosenbeck Gøeg, CLAAUDIA, ITS, AAU

With support from

• DeiC (The Danish e-Infrastructure consortium)
• Aalborg University
• University of Southern Denmark
• Aarhus University
• Center for Humanities Computing

Citation

Andersen, Nikolaj (CLAAUDIA); Gøeg, Pelle Rosenbeck (CLAAUDIA) and OpenAI. (2024). Transcriber (Version1.0) [App]. UCloud interactive HPC system, eScience Center at the University of Southern Denmark. https://cloud.sdu.dk/app/jobs/create?app=transcriber&version=1.0