Annotator

Attention

The GUI annotator is under development and is currently pre-alpha. Use at your own risk and please use version control or back up any critical data.

Currently the functionality of the Annotator GUI allows for users to modify transcripts and add/change entries in the pronunciation dictionary to interactively fix out of vocabulary issues.

Warning

If you are trying to use the annotator from Windows, note that some issues will be present as native Windows use is not fully supported. Specifically if you need G2P functionality, that does not function on Windows due to its dependencies not being available (Pynini, Opengrm-ngram, OpenFst).

To use the annotator, first follow the instructions in Installation. Once MFA is installed and thirdparty binaries have been downloaded, run the following command:

mfa annotator

Initial setup

To load a corpus for inspection, go to the Corpus drop down menu and select “Load a corpus”. Navigate to the desired corpus directory. Please note that it should follow one of the data formats outlined in Data formats.

Note

Some set up of system codecs may be necessary to playback those types of files. For Windows, LAV filters has been tested to work with .flac files.

Next, dictionary files and G2P models should be loaded via their respective menus. If any pretrained models have been installed via Pretrained models, these can be selected directly.

Fixing out of vocabulary issues

Once the corpus is loaded with a dictionary, utterances in the corpus will be parsed for whether they contain an out of vocabulary (OOV) word. If they do, they will be marked in that column on the left with a red cell (see number 2 below).

To fix a transcript, click on the utterance in the table. This will bring up a detail view of the utterance, with a waveform window above and the transcript in the text field. Clicking the Play button (or Tab by default) will allow you to listen to the audio. Pressing the Save current file button (see number 10 below) will save the utterance text to the .lab/.txt file or update the interval in the TextGrid.

Warning

Clicking Save will overwrite the source file loaded, so use this software with caution. Backing up your data and/or using version control is recommended to ensure that any data loss during corpus creation is minimized.

If the word causing the OOV warning is in fact a word you would like aligned, you can right click on the word and select Add pronunciation for 'X' if a G2P model is loaded (see number 7 below). This will run the G2P model to generate a pronunciation in the dictionary which can then be modified if necessary and the dictionary can be saved via the Save dictionary button. You can also look up any word in the pronunciation dictionary by right clicking and selecting Look up 'X' in dictionary. Any pronunciation can be modified and saved. The Reset dictionary button wil discard any changes made to the dictionary.

Fixing segments

Image cannot be displayed in your browser

The file you want to fix up can be selected via the dropdown in the top left (number 1 above).

For fixing up intervals, you can select segments in the left table (number 2 above), or by clicking on intervals in the plot window (i.e., number 5 above). You can edit the text in the center bottom box (number 6 above), change the speaker via the dropdown next to the text box (number 12 below), and adjust boundaries as necessary (green lines associated with number 4 below). If you would like to add a new speaker, then it can be accessed via the Speaker tab on the right pane, which will also list counts of utterances (see 13 below). Entering a speaker name and clicking “Add speaker” (14 below), will make that speaker available in the dropdown.

Single segments can be split via a keyboard shortcut (by default Ctrl+S, but this can be changed, see Configuring the annotator for more details). This will create two segments from one, split at the midpoint, but with all the text in the first segment.

Multiple segments can be selected by holding Ctrl (with selections shown in the left pane, though not in the waveform panel), and can be merged into single segments via a keyboard shortcut (by default Ctrl+M, but this can be changed, see Configuring the annotator for more details). Any number of segments can be selected this way, and the resulting merged segment will concatenate the transcriptions for them all. In general, be cautious about creating too long of utterances, as in general there is better performance in alignment for shorter utterances, and often breath pauses make for good segment boundaries if they’re visible on the waveform.

Image cannot be displayed in your browser

Segments can be added via double clicking on a speaker’s tier (i.e., number 11), however, it is disabled if a segment exists at that point. Any segments can also be deleted via a shortcut (by default Delete). There is limited restore functionality for deleted utterances, via a button on the bottom left.

Configuring the annotator

By going to Preferences in the Edit menu, many aspects of the interface can be changed. The two primary customizations currently implemented are for the appearance of the waveform/segment window and for keyboard shortcuts.

The current available shortcuts are:

Function Default keybind
Play audio Tab
Zoom in Ctrl+I
Zoom out Ctrl+O
Pan left Left arrow
Pan right Right arrow
Merge utterances Ctrl+M
Split utterances Ctrl+S
Delete utterances Del
Save current file By default not bound, but can be set
Create new segment Double click (currently not rebindable)