Section: New Software and Platforms

Speech processing tools

Participants : Denis Jouvet, Dominique Fohr, Odile Mella, Irina Illina, Emmanuel Vincent, Antoine Liutkus, Vincent Colotte, Yann Salaün, Antoine Chemardin.

These automatic speech processing tools deal with audio data transcription (ANTS), audio sources separation (FASST), speech-text alignment (LASTAS) and text-to-speech synthesis (SoJA).

ANTS (Automatic News Transcription System)

ANTS is a multipass system for transcribing audio data, and in particular radio or TV shows. The audio stream is first split into homogeneous segments of a manageable size, and then each segment is decoded using the most adequate acoustic model with a large vocabulary continuous speech recognition engine (Julius or Sphinx). Further processing passes are run in order to apply unsupervised adaptation processes on the features (VTLN: Vocal Tract Length Normalization) and/or on the model parameters (MLLR: Maximum Likelihood Linear Regression), or to use Speaker Adaptive Training (SAT) based models. Moreover decoding results of several systems can be efficiently combined for improved decoding performance. The latest version takes advantage of the multiple CPUs available on a computer, and runs on both standalone linux machines and on clusters.

FASST (Flexible Audio Source Separation Toolbox)

FASST (http://bass-db.gforge.inria.fr/fasst/ ) is a toolbox for audio source separation distributed under the Q Public License. Version 2 in C++ has been developed in the context of the ADT FASST (conducted by MULTISPEECH in collaboration with the PANAMA and TEXMEX teams from Inria Rennes - cf. 8.1.6 ) and released in January 2014. Its unique feature is the possibility for users to specify easily a suitable algorithm for their use case thanks to the general modeling and estimation framework proposed in [6] . It forms the basis of most of our current research in audio source separation, some results of which will be incorporated into future versions of the software.

KAM (Kernel Additive Modelling)

The Kernel Additive Modelling framework for source separation [13] , [42] has been proposed this year by Liutkus et al. as a new and effective approach to source separation. In 2014, two different implementations of KAM have been registered with the APP: a Matlab version matKAM and a python version pyKAM. The former is under a aGPL license, while the latter is under a proprietary license. The rationale for this choice is that the Matlab version is to be mainly disseminated for research purpose to the colleagues in the field, that mainly use Matlab, while the python version is more liable to lead to industrial transferts.

LASTAS (Loria Automatic Speech-Text Alignment Software)

LASTAS is a software for aligning a speech signal with its corresponding orthographic transcription. Using a phonetic lexicon and automatic grapheme-to-phoneme converters, all the potential sequences of phones corresponding to the text are generated. Then, using acoustic models, the tool finds the best phone sequence and provides together the boundaries at the phone level and at the word level.

This year, this software has been included in a web application for speech-text automatic alignement, named ASTALI, which will soon be available (http://astali.loria.fr ).

CoALT (Comparing Automatic Labeling Tools)

CoALT is a software for comparing the results of several automatic labeling processes through user defined criteria  [70] .

SoJA (Speech synthesis platform in Java)

SOJA (http://soja-tts.loria.fr ) is a software for Text-To-Speech synthesis (TTS) which relies on a non uniform unit selection algorithm. It performs all steps from text to speech signal output. Moreover, a set of associated tools is available for elaborating 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 in Java; some are in C. The software runs under Windows and Linux. It can be launched with a graphical user interface or directly integrated in a Java code or by following the client-server paradigm. We will consider extending and making SoJA more modular and able to handle both acoustic and visual features, in order to use it for both acoustic-only synthesis and audiovisual synthesis. In the future, the text-to-speech synthesis platform will get extended to take into account expressivity features.