Section: Software

Software

WinSnoori

Snorri is a speech analysis software that we have been developing for 15 years. It is intended to facilitate the work of the scientist in automatic speech recognition, phonetics or speech signal processing. Basic functions of Snorri enable several types of spectrograms to be calculated and the fine edition of speech signals (cut, paste, and a number of filters) as the spectrogram allows the acoustical consequences of all the modifications to be evaluated. Beside this set of basic functions, there are various functionalities to annotate phonetically or orthographically speech files, to extract fundamental frequency, to pilot the Klatt synthesizer and to utilize PSOLA resynthesis.

The main improvement concerns automatic formant tracking which is now available with other tools for copy synthesis. It is now possible to determine parameters for the formant synthesizer of Klatt quite automatically. The first step is formant tracking, then the determination of F0 parameters and finally the adjustment of formant amplitudes for the parallel branch of the Klatt synthesizer enable a synthetic speech signal to be generated. The automatic formant tracking that has been implemented is an improved version of the concurrent curve formant tracking [55] . One key point of this tracking algorithm is the construction of initial rough estimates of formant trajectories. The previous algorithm used a mobile average applied onto LPC roots. The window is sufficiently large (200 ms) to remove fast varying variations due to the detection of spurious roots. The counterpart of this long duration is that the mobile average prevents formants fairly far from the mobile average to be kept. This is particularly sensitive in the case of F2 which presents low frequency values for back vowels. A simple algorithm to detect back vowels from the overall spectral shape and particularly energy levels has been added in order to keep extreme values of F2 which are relevant.

Together with other improvements reported during the last years, formant tracking enables copy synthesis. The current version of WinSnoori is available on http://www.winsnoori.fr .

LABCORP

contacts : David Langlois (langlois@loria.fr) and Kamel Smaïli (smaili@loria.fr).

In the past, we developed a labelling tool which allows syntactic ambiguities to be solved. The syntactic class of each word is assigned depending on its effective context. This tool is based on a large dictionary (230000 lemmas) extracted from BDLEX and a set of 230 classes determined by hand. This tool has a labelling error of about 1 %.

Such a tool is dedicated to tag a text with predefined set of Parts of Speech . A tagger needs a time-consuming manual pre-tagging to bootstrap the training parameters. It is then difficult to test numerous tag sets as needed for our research activities. However, this stage could be skipped [54] . That's why we developed another tagger based on a unsupervised tagging algorithm. This method has been used to estimate the parameters of a new tagger using the classes of the former one. The new tagger is now integrated into the TTS platform developed in the team (see  5.1.11 ).

Automatic lexical clustering

contacts : David Langlois (langlois@loria.fr) and Kamel Smaïli (smaili@loria.fr).

In order to adapt language models in ASR applications, we use a toolkit we developed in past. This tool automatically creates word classes. This toolkit exploits the simulated annealing algorithm. Creating these classes requires a vocabulary (set of words) and a training corpus. The resulting set of classes minimizes the perplexity of the corresponding language model. Several options are available: the user can fix the resulting number of classes, the initial classification, the value of the final perplexity, etc.

SUBWEB

contacts : David Langlois (langlois@loria.fr) and Kamel Smaïli (smaili@loria.fr).

We published in 2007 a method which allows to align sub-titles comparable copora [57] . In 2009, we proposed an alignment web tool based on the developed algorithm. It allows to: upload a source and a target files, obtain an alignment at a sub-title level with a verbose option, and and a graphical representation of the course of the algorithm. This work has been supported by CPER/TALC/SUBWEB(http://wikitalc.loria.fr/dokuwiki/doku.php?id=operations:subweb ).

ESPERE

contact : Dominique Fohr (fohr@loria.fr).

ESPERE (Engine for SPEech REcognition) is an HMM-based toolbox for speech recognition which is composed of three processing stages: an acoustic front-end, a training module and a recognition engine. The acoustic front-end is based on MFCC parameters: the user can customize the parameters of the filterbank and the analyzing window.

The training module uses Baum-Welch re-estimation algorithm with continuous densities. The user can define the topology of the HMM models. The modeled units can be words, phones or triphones and can be trained using either an isolated training or an embedded training.

The recognition engine implements a one-pass time-synchronization algorithm using the lexicon of the application and a grammar. The structure of the lexicon allows the user to give several pronunciations per word. The grammar may be word-pair or bigram.

ESPERE contains more than 20000 C++ lines and runs on PC-Linux or PC-Windows.

SELORIA

contact : Odile Mella (Odile.Mella@loria.fr).

SELORIA is a toolbox for speaker diarization.

The system contains the following steps:

• Speaker change detection: to find points in the audio stream which are candidates for speaker change points, a distance is computed between two Gaussian modeling data of two adjacent given-length windows. By sliding both windows on the whole audio stream, a distance curve is obtained. A peak in this curve is thus considered as a speaker change point.

• Segment recombination: too many speaker turn points detected during the previous step results in a lot of false alarms. A segment recombination using BIC is needed to recombine adjacent segments uttered by the same speaker.

• Speaker clustering: in this step, speech segments of the same speaker are clustered. Top-down clustering techniques or bottom-up hierarchical clustering techniques using BIC can be used.

• Viterbi re-segmentation: the previous clustering step provides enough data for every speaker to estimate multi-gaussian speaker models. These models are used by a Viterbi algorithm to refine the boundaries between speakers.

• Second speaker clustering step (called cluster recombination): This step uses Universal Background Models (UBM) and the Normalized Cross Likelihood Ratio (NCLR) measure.

This toolbox is derived from mClust designed by LIUM.

ANTS

contact : Dominique Fohr (fohr@loria.fr).

The aim of the Automatic News Transcription System (ANTS) is to transcribe radio broadcast news. ANTS is composed of four stages: broad-band/narrow-band speech segmentation, speech/music classification, detection of silences and breathing segments and large vocabulary speech recognition. The three first stages split the audio stream into homogeneous segments with a manageable size and allow the use of specific algorithms or models according to the nature of the segment.

Speech recognition is based on the Julius engine and operates in two passes: in the first pass, a frame-synchronous beam search algorithm is applied on a tree-structured lexicon assigned with bigram language model probabilities. The output of this pass is a word-lattice. In the second pass, a stack decoding algorithm using a trigram language model gives the N-best recognition sentences.

A real time version of ANTS has been developped. The transcription is done in real time on a quad-core PC.

JSynATS

contact : Christophe Cerisara (Christophe.Cerisara@loria.fr).

JSynATS is the “Java Syntactic parser of Automatically Transcribed Speech”. Its development has started in june 2009 from the collaboration between Parole and Talaris in the context of the RAPSODIS project. It is an open-source dependency parser that is dedicated to oral speech. It departs from the other existing French syntactic parsers from the fact that it aims at efficiently handling transcription errors produced by every automatic transcription systems. JSynATS will participate to the Passage evaluation campaign in november 2009.

JTRANS

contact : Christophe Cerisara (Christophe.Cerisara@loria.fr).

JTrans is an open-source software for semi-automatic alignement of speech and textual corpus. It is written 100% in JAVA and exploits libraries developed since several years in our team. Two algorithms are available for automatic alignment: a block-viterbi and standard forced-alignement Viterbi. The latter is used when manual anchors are defined, while the former is used for long audio files that do not fit in memory. It is designed to be intuitive and easy to use, with a focus on GUI design. The rationale behind JTrans is to let the user control and check on-the-fly the automatic alignment algorithms. It is bundled for now with a French phonetic lexicon and French models, but an English version may be released in the future.

JTrans is developed in the context of the CPER MISN TALC project, in collaboration between the Parole and Talaris INRIA teams, and CNRS researchers from the ATILF laboratory. It is distributed under the Cecill-C licence, and can be downloaded at http://jtrans.gforge.inria.fr

STARAP

contact : Dominique Fohr (fohr@loria.fr).

STARAP (Sous-Titrage Aidé par la Reconnaissance Automatique de la Parole) is a toolkit to help the making of sub-titles for TV shows. This toolkit performs:

• Parameterization of speech data;

• Clustering of parameterized data;

• Gaussian Mixture Models (GMM) training;

• Viterbi recognition.

This toolkit was realised in the framework of the STORECO contract and the formats of the input and output files are compatible with HTK toolkit.

TTS SoJA

contact : Vincent Colotte (Vincent.Colotte@loria.fr).

TTS SoJA (Speech synthesis platform in Java) is a software of text-to-speech synthesis system. The aim of this software is to provide a toolkit to test some steps of natural language processing and to provide a whole system of TTS based on non uniform unit selection algorithm. The software performs all steps from text to the speech signal. Moreover, it provides a set of tools to elaborate a corpus for a TTS system (transcription alignment, ... ). Currently, the corpus contains 1800 sentences (about 3 hours of speech) recorded by a female speaker.

Most of the modules are developed in Java. Some modules are in C. The platform is designed to make easy the addition of new modules. The software runs under Windows and Linux (tested on Mandriva, Ubuntu). It can be launch with a graphical user interface or directly integrated in a Java code or by following the client-server paradigm.

The software license should easily allow associations of impaired people to use the software.