This tool can load an output of a keyword-spotting system (KWS) and reference file in HTK-MLF format and show detections in a tabular view. You can also use it to replay detections, tune and visualize scores, hits, misses and false-alarms using sliders on the right-side panel.
Download here kwsviewer_v1.5_win32.zip (6MB). No installation is needed. You can run kwsviewer.exe directly without any installation.
Download here kwsviewer_v1.5_src.zip (300kB). Requirements:
Unpack the kwsviewer's source archive, edit paths in kwsviewer.pro, run "make" and you should have the executable ./kwsviewer ready.
First you have to open an MLF (output of the keyword-spotting system). In the main menu go to "File" -> "Open" and select the MLF with detected keywords. You can also set a reference MLF to see hits, misses and false alarms. If you would like to listen to the detections, set also the "Audio folder". After the MLF is loaded, you can:
Each detection in the table view has a green or red background color depending on it's Normalized 0..1 score. The closer the score is to 1, the greener is the color and the further it is to 0, the redder is the color. The threshold between green and red is set to 0.5.
To open the filter dialog, go to "Settings" -> "Keywords filter" or press "Ctrl+f". Here you can select the keywords which you want to see in the table view.
You can also double-click on any item in any of the two lists to move it to the other list.
To open the preferences dialog, go to "Settings" -> "Preferences" or press "Ctrl+p". You can set there:
Each detection belongs to some file (the "Filename" column). If the audio directory was set, then the filename with the extension specified in the "Preferences" dialog is searched in the audio directory. If it is not found there, media player will take the detection's file with it's full path and replace it's extension (after the last dot) with the specified audio file extension. If you double-click on any detection, the part of the audio file given by the detection's start and end time will be replayed. You can expand the time boundaries in the Preferences dialog -> "Detection context for audio player". You can also use the media player controls on the right-side panel:
You can also use the slider to seek to any position in the audio file.
- multiplicative_constant ...additive and multiplicative constant is set on the right-side panelof the application.
29600000 39200000 complicated 0.0000000000 0.9998064924 1
3100000 9700000 usability -0.0000000000 0.9998064924 1
27800000 33500000 train -3.8202736792 0.4460323491 1
36700000 41700000 file -0.0060353995 0.9320198154 1
Where the meaning of columns is:
1st - start time of the detection in hundreds of nanoseconds (time_in_seconds * 10000000)
2nd - end time of the detection in the same time units
3rd - label of the detection
4th - score of the detection (either log-probability or log-likelihood-ratio)
5th - normalized score of the detection
6th - hard decision
Only the first four columns are required as the output of keyword spotter, the others are optional. For the reference file, only first three columns are needed.