The RWTH Aachen LVCSR system for IWSLT-2016 German Skype conversation recognition task

The RWTH Aachen LVCSR system for IWSLT-2016 German Skype conversation recognition task Wilfried Michel, Zolt´an T¨uske, M. Ali Basha Shaik, Ralf Schl¨uter, Hermann Ney Human Language Technology and Pattern Recognition – Computer Science Department RWTH Aachen University, 52056 Aachen, Germany { michel, tuske, shaik, schlueter, ney }@cs.rwth-aachen.de Abstract In this paper the RWTH large vocabulary continuous speech recognition (LVCSR) systems developed for the IWSLT- 2016 evaluation campaign are described. This evaluation campaign focuses on transcribing spontaneous speech from Skype recordings. State-of-the-art bidirectional long short- term memory (LSTM) and deep, multilingually boosted feed-forward neural network (FFNN) acoustic models are trained an narrow and broadband features. An open vocabu- lary approach using subword units is also considered. LSTM and count-based full word and hybrid backoff language mod- eling methods are used to model the morphological richness of the German language. All these approaches are combined using confusion network combination (CNC) to yield a com- petitive WER. 1. Introduction The International Workshop on Spoken Language Transla- tion (IWSLT) is an annual workshop including an evalu- ation campaign in the tasks of automatic speech recogni- tion (ASR), machine translation (MT), and spoken language translation (SLT), which is the union of the aforementioned. Participants have the opportunity to compete in this evalua- tion campaign to compare the strength of their systems and advance the state of the art. The German ASR task of the 2016 evaluation campaign employs data from the Microsoft Speech Language Transla- tion (MSLT) task. The goal is to transcribe one side of bilin- gual Skype voice calls. Development and test data are audio files with a varying length between 1 and 30 seconds each containing one part of an informal conversation between two natural persons. This is challenging from multiple points of view. Informal conversations tend to be highly spontaneous which leads to an increased number of disfluencies and mis- articulations. Conversational speech also covers a broader range of topics which makes it hard to train and fine tune appropriate language models. To overcome these difficulties we propose a combina- tion of two different acoustic models and two different lan- guage models. We used the LIUM auto-segmenter [1] to trim non-speech parts and to segment the longer of the au- dio files. From the audio files we extracted standard cepstral features: MFCC, PLP, Gammatone [2] and also the critical band energies (CRBE) of the corresponding pipelines. A hy- brid LSTM acoustic model was trained on the Gammatone features to directly output tied-triphone state posterior proba- bilities. To generate a second, largely complementary acous- tic model for German, we trained a Gaussian mixture model based HMM processing multilingually initialized deep bot- tleneck feed-forward NNs using the tandem approach [3, 4]. Both systems are complemented with two 5-gram lan- guage models for initial decoding [5]. In a second pass the generated lattices are rescored using LSTM recurrent NN based language models. In a final step all four lattices are combined using confusion network combination. The re- sults were achieved using RETURNN, the RWTH …

Content truncated to 500 words