Section: Software and Platforms

ANTS

The aim of the Automatic News Transcription System (ANTS) is to transcribe radio or TV shows. ANTS is composed of several stages. The first processing steps aim at splitting the audio stream into homogeneous segments of a manageable size and at identifying the segment characteristics in order to allow the use of specific algorithms or models according to the nature of the segment. This includes broad-band/narrow-band speech segmentation, speech/music classification, speaker segmentation and clustering, detection of silences/breathing segments and generally speaker gender classification.

Each segment is then decoded using a large vocabulary continuous speech recognition engine, either the Julius engine or the Sphinx engine. The Julius engine 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. The Sphinx engine processes the speech input segment in a single forward pass using a trigram language model.

Further processing passes are usually run in order to apply unsupervised adaptation processes on the feature computations (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 both systems can be efficiently combined for improved decoding performance.

The latest version which relies on a perl script exploits the multiple CPUs available on a computer to reduce the processing time, and runs on both a stand alone linux machine and on the cluster.