Using the Training-module, you can improve your acoustic model.

The interface lists all installed training texts in a table with three columns:

  • Name

    A descriptive name for the text.

  • Pages

    The number of pages the text consists of. Each page represents one recording.

  • Recognition Rate

    Analogue to the vocabulary; represents how likely Simon will recognize the words (higher is better). The recognition rate of the training text is the average recognition rate of all the words in the text.

To improve the acoustic model - and thus the recognition rate - you have to record training texts. This means that Simon gets essentially two needed parts:

  • Samples of your speech

  • Transcriptions of those samples

The active dictionary is used to transcribe the words (mapping them from the actual word to its phonetic transcription) that make up the text so every word contained in the training text you want to read (train) has to be contained in your active dictionary. Simon will warn you if this is not the case and provide you with the possibility to add all the missing words in one go.

The procedure is the same as if you would add a single word but the wizard will prompt you for details and recordings for all the missing words automatically. This procedure can be aborted at any time and Simon will provide both a way to add the already completely defined words and to undo all changes done so far. When the user has added all the words he is prompted for (all the words missing) the changes to the active dictionary / vocabulary are saved and the training of the previously selected text starts automatically.

The training (reading) of the training text works exactly the same as the initial training when adding a new word.

Make sure you follow the guidelines listed in the recording section.

Storage Directories

Training texts are stored in two different locations:

  • Linux: ~/.kde/share/apps/simon/texts

    Windows: %appdata%\.kde\share\apps\simon\texts

    The texts of the current user. Can be deleted and added with Simon (see below).

  • Linux: `kde4-config --prefix`/share/apps/simon/texts

    Windows: (install folder)\share\apps\simon\texts

    System-wide texts. They will appear on every user account using Simon on this machine and cannot be deleted from within Simon because of the obvious permission restrictions on system-wide files.

    This folder can be used by system administrators to provide a common set of training texts for all the users on one system.

The XML files (one for each text) can just be moved from one location to the other but this will most likely require admin privileges.

Adding Texts

The add texts wizard provides a simple way to add new training texts to Simon.

When importing text files, Simon will automatically try to recognize individual sentences and split the text into appropriate pages (recordings). The algorithm treats text between normal punctuation (., !, ?, ..., ",...) and line breaks as sentences. Each sentence will be on its own page.

Simon supports two different sources for new training texts.

Add training texts

Simply enter the training text in an input field.

Local text files

Simon can import normal text files to use them as training texts.

On-The-Fly Training

In addition to training texts, Simon also allows to train individual words or word combinations from your dictionary on-the-fly.

This feature is located in the vocabulary menu of Simon.

Select the words to train from the vocabulary on the left and simply drag them to the selection list to the right (you could also select them in the table on the left and add them by clicking Add to Training).

Start the training by selecting Train selected words. The training itself is exactly the same as if it were a pre-composed training text.

If there are more than 9 words to train Simon will automatically split the text evenly across multiple pages.

Of course you are free to add words from the shadow lexicon to the list of words to train but Simon will prompt you to add the words before the training starts just like he would if you would train a text that contains unknown words (see above).