A simple tool to help transcribe audio clips for machine learning datasets
Specficically, it was designed to make datasets for Tacotron 2 due to it's popularity with the uberduck.ai community, so all the exported files and warnings are associated with making a good dataset for it.
You can always download the latest version from the Releases Tab
You will need to load a folder containing all your wav files. You can do this via drag-and-drop:
You will see some stats about your audio, and a list containing all your audio files, with textboxes for you to transcribe the speech. Anything you type is automatically saved to a .txt file with the same name as your .wav file. Autosaves happen 1 second after you stop typing, as well as when you click away from the textbox.
If your audio files have problems, you will see a red triangle with an exclamation mark next to it. You can hover over the error symbol to get a list of all the problems your audio file has:
Note: You will have to fix the audio files yourself, and there is currently no way to "reload" the files. You will need to restart the application to load the audio files again.
If your transcript has problems, the characters will be marked red, and a tooltip will appear when you hover over the textbox, explaining what is wrong:
Once you're finished, simply hit the "Export transcript" button at the bottom of the page, and a file called transcript.txt will be created in the folder with all your wavs.
You can then use this file in the tacotron training notebook.
If you want to build the app yourself, simply download the repo using
git clone https://github.com/underscoren/transcribetool.git
Then, install all the dependencies using
npm install
Finally, build the app using
npm run make
The app will build into the out folder, usually as both the application files and an installer you can use for your system.
- Manual reload of audio files
- Automatic detection when audio files are changed
- Editing/Auto-fixing/Deleting audio clips
- Custom export location
- Icons
The source code for this application is licensed under MIT and all the various libraries used contain their own licenses in their respective files.