User Guide

A detailed description of ArBade functionalities is discussed in this document.

1. Record
   • Record Status
   • Record Parameters
   • Record Shortcuts
2. Verify
   • Verify Status
   • Verify Parameters
   • Verify Shortcuts
3. Sleep
4. Wrong
5. False
6. Test
7. Enn False
8. Console

Record

While recording the samples for training the voice recognition engine, you should know the status, parameters and shortcuts related to recording.

Record Status

• Rec: Records your voice
• Pause: Pause the Recording Procedure
• Stop: Stop the record and navigate to statistics panel.
• Req Pause: Requested pause will be commited after recording finished
• Break: Time to let you read the words before recording started

Record Parameters

• Category: Directories to classify samples you record with different devices. exceptions are unverified, online, sleep, wrong, test, enn and efalse which are used internally by Benjamin.
• Pause Time: Break time that lets you read the words before record process starts.
• Num of Words: Specify the number of words for recording samples.
• Rec Time: Defines the period of recording time.
• Count: Set this to the total number of samples you want to record in a single round.
• Focus Word: A displayed set of words will contain the Focus Word.
• Power: After recording the sample power of the voice will be displayed. Care about this parameter as this will be very low if the recording device has any problem.
• Time: Shows how much time passed from Rec Time
• Word: Shows the words you should say while recording
• Status: Shows recording status

Record Shortcuts

• S: Set Category, unverified and efalse are prohibited. A dialog will be shown and ask for the desired Category. You can create a new Category or change between categories produced in the past.
• Space: Start or Pause recording.
• Up/Down: Increase/Decrease Pause Time.
• Right/Left: Increase/Decrease Rec Time.
• C: Change count, a dialog will be opened and ask for how many samples would you record.
• F: Set Focus Word, then the displayed words collection will contain Focus Word. A dialog will be opened and ask for the id of Focus Word.
• W: Opens a dialog to ask Num of Words.
• O: Opens selected Category directory.

Verify

By using Benjamin regularly, many samples will be created as online samples. However, these samples are not labeled correctly. There may be some mistakes among words from a sample. These samples are accumulated in unverified directory and by verifying them, they will be moved to online directory.

Verification is done by playing the samples in unverified directory. you can refer to this section to know all status, parameters and shortcuts related to verification.

Verify Shortcuts

• Space: Start or Pause playing.
• Up/Down: Increase/Decrease Pause Time.
• F: Set Focus Word, then the displayed words collection will contain Focus Word. A dialog will be opened and ask for the id of Focus Word.
• O: Opens unverified directory.
• Z: Press to move the sample from the unverified directory to the wrong directory.
• R: Changes default decision/action for verifying samples. Copy mode means after the Decide Pause timeout, the sample will be copied to the online directory. In the same way in Delete mode, timeout will cause sample removal.

Sleep

Verify samples recorded in sleep mode. In this mode listen to the recorded samples, if they contain any true samples press Z to get rid of them. All other samples will automatically moved to the wrong directory and they will be used to train the enn model.

Wrong

These are verified samples from the sleep directory. They are verified to be wrong samples, not true ones! In this mode, you will be able to review them and check if some samples are moved to this directory by accident.

False

During generating Enn samples, some of the samples in the train directory will not detect correctly, these samples will be moved to the efalse directory. In this mode, you can listen to these samples and judge whether they are real samples or the wrong ones which you should get rid of them.

Test

After training on samples, the model is evaluated through test samples. Word error rate (WER) and Sentence error rate (SER) are calculated by predicting test samples. Then false detected samples are listed in this tab and you can listen to them to figure out what went wrong.

Enn False

BaTool generates enn samples, from both train samples and wrong samples. The enn directory is the destination of this process. Enn files from the train directory are located in enn/true and wrong directory-related files are in enn/false. In the Enn False tab, only false statistics are shown. By clicking the trash icon, all Enn samples will be deleted.

Console

Training on samples is started by pressing T. After training is finished, a dialog pops up and asks to generate enn samples from audio samples. Neural Network will be trained from these enn samples. When generating Enn Samples is finished, another dialog arises and asks about verifying generated engine false detected samples. Engine false detected samples, abbreviated as efalse samples, are collected in the efalse directory. They are generated after training, and while testing samples in the test directory. You can verify them by switching to False and Test tabs.