Voice Import Tools Tutorial : How to build a new Voice with Voice Import Tools

This Tutorial explains the procedure to build a new voice with Voice Import Tools (VIT) under MARY Environment.

Voice Import Tool is a Graphical User Interface(GUI), which contains a set of Voice Import Components and helps the user to build new voices under MARY(Modular Architecture for Research in speech sYnthesis) Environment. This GUI Tool designing is primarily aims to build new voices very easily by any user with out knowing much technical details of Speech Synthesis.

Currently, Voice Import Tool supports following categories mainly:

1. Feature Extraction from Acoustic Data
2. Feature Vector Extraction from Text Data
3. Automatic Labeling
4. Unit Selection
5. Voice Installation to MARY

Requirements Needed:

• Operating System - Linux (Recommended)
• MARY TTS Recent Version - Download Link: ​http://mary.dfki.de/Download
• Openmary - SVN from ​http://mary.opendfki.de

(we also able to use Windows also, if we can able to compile properly the following dependent tools.)

Dependendent Tools:

• Praat Pitch Marker or Snack - For pitch marks

Download Link for praat : ​http://www.fon.hum.uva.nl/praat

Installing Snack : Require tcl and snack. Installation instructions available at ​http://www.speech.kth.se/snack/

• Edinburgh Speech Tools Library – For MFCCs and Wagon (CART)

Download Link for Speech Tools: ​http://www.cstr.ed.ac.uk/projects/speech_tools/

• EHMM or Sphinx – For Automatic Labeling

EHMM is available with festvox-2.1 (Recent Version) - ​http://festvox.org/download.html

Sphinx - ​http://cmusphinx.sourceforge.net/webpage/html/download.php

Voice Import Components:

Following Components are available with Voice Import Components:

• PraatPitchmarker
• SnackPitchmarker
• MCEPMaker
• Festvox2MaryTranscripts
• Mary2FestvoxTranscripts
• PhoneUnitFeatureComputer
• HalfPhoneUnitFeatureComputer
• EHMMLabeler
• SphinxLabelingPreparator
• SphinxTrainer
• SphinxLabeler
• MRPALabelConverter
• LabelledFilesInspector
• PhoneUnitLabelComputer
• PhoneLabelFeatureAligner
• HalfPhoneUnitLabelComputer
• HalfPhoneLabelFeatureAligner
• HalfPhoneUnitfileWriter
• HalfPhoneFeatureFileWriter
• JoinCostFileMaker
• AcousticFeatureFileWriter
• CARTBuilder
• CARTPruner
• VoiceInstaller

How to run?

1. First you need to have following 2 basic requirements for Voice Building

1. Wave files
2. Corresponding Transcription (in MARY or Festival Format)

MARY Format : Each transcription represented by a single file. All these files placed in a single directory. By default, all these files placed in 'text' directory of voice-building directory.

Festival Format : A single file contains all transcriptions. For examples see below example.

( arctic_a0001 "AUTHOR OF THE DANGER TRAIL, PHILIP STEELS, ETC" )
( arctic_a0002 "Not at this particular case, Tom, apologized Whittemore." )
( arctic_a0003 "For the twentieth time that evening the two men shook hands." )
( arctic_a0004 "Lord, but I'm glad to see you again, Phil." )
( arctic_a0005 "Will we ever forget it." )

2. Create a new Voice Building Directory

• Put all Wave files in "wav" directory

3. Run below commands through Shell script from Voice Building Directory.

export MARY_BASE="/path/to/mary"
java -jar -Xmx1024m  $MARY_BASE/java/voiceimport.jar

GUI is looking like below (Which supports voice building):

When you are running first time above shell script, It asks you some basic configuration settings by presenting with a GUI window where you have to enter a few basic settings. Almost all other settings are based on these first settings and set automatically. After clicking the "Save"-button, you get to the main window. There you can see a list of modules. A component is executed by ticking the associated checkbox and clicking on "Run".

Global Configuration Settings window looks like below:

Global Configuration Settings:

Domain - general or limited

Gender - male or female

Locale - which specifies language of domain (de - Deutsch or en - English)

(Currently, MARY supporting 2 language only: 1. Deutsch 2. English)

Marybase - MARY Installation Directory (Global Path)

Rootdir - Voice Building Directory (Global Path)

Wavdir - Where we can store Wave files

Textdir - Where we can store corresponding Transcriptions

User also can change the settings for each individual component by clicking on the wrench symbol next to the component. Clicking on "Settings" takes you to the window where you can change the basic settings. In a settings window, you can change the view to the settings of another module or the basic settings via the drop-down menu. Basically, all modules need to be run to import the voice into MARY. For more detailed information, check the general help file - just click on "Help" in the main window. Clicking on help in the settings window opens a help window with details about the displayed settings. We recommended to give Absolute Paths for individual Configuration Settings. These config. settings are arguments to components to perform corresponding task.

The import tool creates two files in the directory where you started it - database.config and importMain.config. database.config contains the values of the settings - you can change the settings also in this file, but be aware that this may cause problems.

Simplest way of Using Voice Import Components:

• Give Config. Settings for Each and Every Component.
• Tick mark all components
• Click RUN button
  

It can complete all tasks in sequential manner.

But No need to use all components for Building a New Voice. For Example: For Automatic Labeling we can choose EHMM or Sphinx.

Explanation on Individual Voice Import Components

1. Feature Extraction from Acoustic Data

PraatPitchmarker

It computes pitch markers with help of Praat. You need to compile or install Praat in your machine.

It also do corrections for Pitch Marks to align near by Zero Crossing. Configuration Settings:

• command - Give Absolute path of Praat Executable
• pmDir - Output Dir Path for Praat Pitch marks
• corrPmDir - Output Dir Path for corrected pitch marks (Pitch marks tuned towards Zero Crossing)
• maxPitch, minPitch - For choosing Pitch Range (Ex: Male: 50-200 | Female: 150-300)

MCEPMaker

It calculate MFCCs from Speech Wave files, using Edinburgh Speech Tools.

Configuration Settings:

• estDir - Edinburgh Speech Tools Compiled Directory
• pmDir - Praat Pitch marks Directory
• corrPmDir - Corrected Pitch marks Directory
• mcepDir - Output Dir for MFCCs

2. Support for Transcription Conversion

Festvox2MaryTranscripts

This Component supports user to convert Festvox Transcription format (ex: txt.done.data) to MARY Supportable format. MARY contains individual text files for each wave file. All Voice Import Components use Transcription from MARY Format. So This component is very useful, if user have Transcription in Festvox format.

Configuration Settings:

• transcriptFile - Festvox format transcription file (Absolute path)

Mary2FestvoxTranscripts

It supports user to convert MARY Supportable format to Festvox format Transcription. It does reciprocal process to above component.

Configuration Settings:

• transcriptFile - Output Festvox format transcription file (Absolute path)

3. Feature Vector Extraction from Text Data

PhoneUnitFeatureComputer

PhoneUnitFeatureComputer computes Phone feature vectors for Unit Selection Voice building process.

• Note: This module requires a running Maryserver from MARY Installation.
  

You can connect to a different server by altering the settings. See the settings help for more information on this. What type of features computed is depends on configuration file called "targetfeatures.config". This configuration file is in Marybase/conf/ directory and directs Server to compute feature vectors.

Configuration Settings:

• featureDir - Output Directory to place computed Phone feature vectors (Absolute path)
• maryServerHost - Server Name
• maryServerPort - Socket Port number (Default 59125)

HalfPhoneUnitFeatureComputer

This component also same as above component. But It computes Half phone level feature vectors. Here "halfphone-targetfeatures.config" file, which is in Marybase/conf/ directory directs Server to compute Half-Phone level feature vectors.

Configuration Settings:

• featureDir - Output Directory to place computed Half-Phone feature vectors (Absolute path)
• maryServerHost - Server Name
• maryServerPort - Socket Port number (Default 59125)

4. Automatic Labeling

EHMMLabeler

EHMM Labeler is a labeling tool, which generates label files with help of Wave files and corresponding Transcriptions. EHMM basic tool is available with Festvox Recent Version. For running EHMM Labeler under MARY environment you need to compile EHMM tool in your machine. It may take long time depending on the size of the data and system configuration.

EHMMLabeler Supports:

1. Database labeling with Force alignment by Training with Flat-Start Initialization
2. Database Labeling with Force alignment by Training with initialized models (Re-Training)
3. Database Labeling with Force alignment by already existed models (Decoding only)

Configuration Settings:

• ehmmDir - EHMM basic package compilation Directory.
• eDir - Directory name (Absolute path) to copy Transcription (in ehmm Supported format) and to store ehmm model.
• featureDir - Feature vectors Directory path, where phone features vectors were computed. (To get phone sequence)
• startEhmmModelDir - Already existing EHMM model Directory path to Initialize EHMM models (for Re-training or Decoding)
• reTrainFlag - (true | false) true - Do re-training by initializing with given models. false - Do just Decoding
• outputLabDir - Dir. Path to store generated Labels

Automatic Labeling using Sphinx Tools:

SphinxLabelingPreparator, SphinxTrainer and SphinxLabeler Components used to do Automatic Labeling with Sphinx tools. These 3 components need SphinxTrain, Sphinx Decoder and Edinburgh Speech Tools for training models and Force alignment.

SphinxLabelingPreparator

This Component prepares the required setup needed for SphinxTrain to train Models.

Configuration Settings:

• estDir - Edinburgh Speech Tools Compiled Directory
• maryServerHost - Server Name
• maryServerPort - Socket Port number (Default 59125)
• sphinxTrainDir - SphinxTrain installation Directory
• stDir - Directory name (Absolute path) to copy Dictionaries and Temp. files (in Sphinx Supported format).
• transcriptFile - Festvox format transcription file (Absolute path)

SphinxTrainer

It trains models required for labeling using Sphinxtrain. It may take long time depending on the size of the data and system configuration.

Configuration Settings:

• stDir - Absolute path of directory where all Dictionaries and Temp. files stored by SphinxLabelingPreparator.

SphinxLabeler

It produces labels with the help of the models built by the SphinxTrainer. It uses Sphinx-2 Decoder for force alignment.

Configuration Settings:

• sphinx2Dir - Sphinx-2 Installation directory absolute path.
• stDir - Absolute path of directory where all Dictionaries, Temp. files and models stored by SphinxLabelingPreparator and SphinxTrainer.

MRPALabelConverter

If you have labeled data in the Festvox format and using the MRPA-Phoneset, use this module to convert the phones into the phoneset used by Mary.

Configuration Settings:

• mrpaLabDir - MRPA Label file directory

5. Label or Pause Correction and Label-Feature Alignment

LabelledFilesInspector

It allows user to browse through aligned labels and listen to the corresponding wave file. It is useful for perceptual manual verification on alignment.

Configuration Settings:

• corrPmDir - Directory Path for corrected pitch marks.

PhoneUnitLabelComputer and HalfPhoneUnitLabelComputer

These components converts the label files into the label files used by Mary. PhoneUnitLabelComputer produces phone labels, HalfPhoneUnitLabelComputer produces halfphone labels. User need both to build the voice.

Configuration Settings:

• labelDir - Output phone label dir. path for PhoneUnitLabelComputer.

Output Half phone label dir. path HalfPhoneUnitLabelComputer.

PhoneLabelFeatureAligner

It tries to align the labels and the feature vectors. If alignment fails, you can start the automatic pause correction.

This works as follows:

• pauses, that are in the label file but not in the feature file are deleted in the label file, and the durations of the previous and next labels are stretched.
• pauses that are in the feature file but not in the label file are inserted into the label file with length zero.

If there are still errors after the pause correction, you are prompted for each error. You can skip the error or remove the corresponding file from the basename list (the list of files that are used for your voice). "skip all" and "remove all" does this for all problematic files. "Edit unit labels" allows you to edit the label file. "Edit RAWMARYXML" lets you edit the maryxml that is the input for computing the features. You have to have a Maryserver running in order to recompute the features from the maryxml. You can alter the host and port settings for the server by altering the settings for the UnitFeatureComputer.

Configaration Settings:

• featureDir - Phone feature vectors directory
• labDir - Phone Labels directory

HalfPhoneLabelFeatureAligner

It also works same as PhoneLabelFeatureAligner, but it works for halfphone units case.

Configuration Settings:

• featureDir - Half Phone feature vectors directory
• labDir - Half Phone Labels directory

6. Basic Data Files

Following components will create basic binary files, which contain whole voice database. So that it is easier and faster to access Database. These files are needed for various voice building steps and for synthesis.

WaveTimelineMaker

The WaveTimelineMaker split the waveforms as datagrams to be stored in a timeline in Mary format. It produces a binary file, which contains all wave files.

Configuration Settings:

• corrPmDir - Directory Path for corrected pitch marks.
• WaveTimeline - file containing all wave files. Will be created by this module

BasenameTimelineMaker

The BasenameTimelineMaker takes a database root directory and a list of basenames, and associates the basenames with absolute times in a timeline in Mary format.

Configuration Settings:

• pmDir - Directory containing the pitchmarks
• timelineFile - file containing the list of files and their times, which will be created by this module.

MCepTimelineMaker

The MCepTimelineMaker takes a database root directory and a list of basenames, and converts the related wav files into a mcep timeline in Mary format.

Configuration Settings:

• mcepDir - directory containing the mcep files
• mcepTimeline - file containing all mcep files. Will be created by this module

7. Building acoustic models

PhoneUnitfileWriter

It produces a file containing all phone sized units.

Configuration Settings:

• corrPmDir - Directory containing the corrected pitchmarks
• labelDir - Directory containing the phone labels
• unitFile - File containing all phone units. Will be created by this module

PhoneFeatureFileWriter

It produces a file containing all the target cost features for the phone sized units. The module needs a file defining which features are to be used and what weights are given to them. They must be the same features as the ones that the PhoneFeatureComputer used. If you do not have a feature definition, the module tries to create one.

For more information, see the example file: Marybase/lib/modules/import/examples/PhoneUnitFeatureDefinition.txt

Configuration Settings:

• featureDir - directory containing the phone features
• featureFile - file containing all phone units and their target cost features.Will be created by this module
• unitFile - file containing all phone units
• weightsFile - file containing the list of phone target cost features, their values and weights

DurationCARTTrainer

It builds an acoustic model of durations in the database using the program "wagon" from the Edinburgh Speech tools.

Configuration Settings:

• durTree - file containing the duration CART. Will be created by this module
• estDir - directory containing the local installation of the Edinburgh Speech Tools
• featureDir - directory containing the phonefeatures
• featureFile - file containing all phone units and their target cost features
• labelDir - directory containing the phone labels
• stepwiseTraining - "false" or "true"
• unitFile - file containing all phone units
• waveTimeline - file containing all wave files

F0CARTTrainer

It builds acoustic models of F0 like DurationCARTTrainer. It uses "wagon" and the files produced by PhoneUnitfileWriter and PhoneFeatureFileWriter.

Configuration Settings:

• estDir - directory containing the local installation of the Edinburgh Speech Tools
• f0LeftTreeFile - file containing the left f0 CART. Will be created by this module
• f0MidTreeFile - file containing the middle f0 CART. Will be created by this module
• f0RightTreeFile - file containing the right f0 CART. Will be created by this module
• featureDir - directory containing the phonefeatures
• featureFile - file containing all phone units and their target cost features
• labelDir - directory containing the phone label files
• stepwiseTraining - "false" or "true"
• unitFile - file containing all phone units
• waveTimeline - file containing all wave files

8. Unit Selection

HalfPhoneUnitfileWriter

It produces a file containing all halfphone sized units.

Configuration Settings:

• corrPmDir - directory containing the corrected pitchmarks
• labelDir - directory containing the halfphone labels
• unitFile - file containing all halfphone units. Will be created by this module

HalfPhoneFeatureFileWriter

It produces a file containing all the target cost features for the phone sized units. The module needs a file defining which features are to be used and what weights are given to them. They must be the same features as the ones that the HalfPhoneFeatureComputer used. If you do not have a feature definition, the module tries to create one.

For more information, see the example file: Marybase/lib/modules/import/examples/HalfPhoneUnitFeatureDefinition.txt

Configuration Settings:

• featureDir - directory containing the halfphone features
• featureFile - file containing all halfphone units and their target cost features.Will be created by this module
• unitFile - file containing all halfphone units
• weightsFile - file containing the list of halfphone target cost features, their values and weights

JoinCostFileMaker

It produces a file containing all the join cost features for the halfphone sized units.

Configuration Settings:

• joinCostFile - file containing all halfphone units and their join cost features. Will be created by this module
• mcepDir - directory containing the mcep files
• mcepTimeline - file containing all mcep files
• unitFile - file containing all halfphone units
• weightsFile - file containing the list of join cost weights and their weights

AcousticFeatureFileWriter

It produces a file containing all the target cost features plus two acoustic target cost features for the halfphone sized units. Also produces a feature definition containing those features.

• acFeatDef - file containing the list of phone target cost features, their values and weights
• acFeatureFile - file containing all halfphone units and their target cost features plus the acoustic target cost features. Will be created by this module.
• featureFile - file containing all halfphone units and their target cost features
• unitFile - file containing all halfphone units
• waveTimeLine - file containing all wave files

CARTBuilder

It builds a preselection tree for the target cost features using "wagon" (CART) from the Edinburgh Speech tools.

Additionally, User need to specify either a feature sequence or a top level tree. They are used to built a basic tree that is extendend by wagon. This way, wagon runs several times on smaller subsets of units rather than the whole set. It might still take some time to run this module.

• Feature sequence: A file containing a list of features for which to build the tree.
• Top level tree: A file containing the basic tree.

For more information on these two possibilities of specifying the basic tree, see the example files in Marybase/lib/modules/import/examples/

If you give the CARTBuilder neither a feature sequence nor a top level tree file, a default feature sequence is created which only contains "mary_phoneme" as feature. If the basic tree contains leaves that are contain more units than the maximum number of units allowed, the leaves are pruned and a warning message is printed. It is recommended that you make sure that there are no leaves that are too big.

Configuration Settings:

• acFeatureFile - file containing all halfphone units and their target cost features plus the acoustic target cost features
• cartFile - file containing the preselection CART. Will be created by this module
• estDir - directory containing the local installation of the Edinburgh Speech Tools
• featureSeqFile - file containing the feature sequence for the basic tree
• maxLeafSize - the maximum number of units in a leaf of the basic tree
• mcepTimeline - file containing the mcep files
• readFeatureSequence - if "true", basic tree is read from feature sequence file; if "false", basic tree is read from top level tree file.
• topLevelTreeFile - file containing the basic tree
• unitFile - file containing all halfphone units

CARTPruner

It prunes the preselection tree and this module also removes outliers from the preselection tree.

Configuration Settings:

• cartFile - file containing the preselection CART
• prunedCartFile - file containing the pruned preselection CART. Will be created by this module
• unitFeatureFile - file containing all halfphone units and their target cost features
• unitFile - file containing all halfphone units
• waveFile - file containing all wave files

9. Installation of New Voice in to MARY

VoiceInstaller

It supports the built voice installation in to MARY automatically. It copies all the necessary files to a new subdirectory in the lib/voices/ directory of your Mary installation. Furthermore, a file that specifies the properties of the voice is created and stored in the conf/ directory of your Mary installation. Next time you start the Mary server, the voice is loaded.

Configuration Settings:

• cartFile - file containing the preselection CART
• durTree - file containing the duration CART
• exampleText - file containing example text (for limited domain voices only)
• f0LeftTree - file containing the left f0 CART
• f0MidTree - file containing the mid f0 CART
• f0RightTree - file containing the right f0 CART
• halfPhoneFeatDefAc - file containing the list of halfphone target cost features, their values and weights
• halfPhoneFeatsAc - file containing all halfphone units and their target cost featuresplus the acoustic target cost features
• halfPhoneUnits - file containing all halfphone units
• joinCostFeatDef - file containing the list of join cost weights and their weights
• joinCostFeats - file containing all halfphone units and their join cost features
• phoneFeatDef - file containing the list of phone target cost features, their values and weights
• waveTimeline - file containing all wave files