RACAI Entry for the IWSLT 2016 Shared Task

RACAI Entry for the IWSLT 2016 Shared Task Sonia Pipa, Alin Florentin Vasile, Ioana Ionașcu, Stefan Daniel Dumitrescu, Tiberiu Boros Research Institute for Artificial Intelligence “Mihai Drăgănescu” Romanian Academy, Bucharest sonia@racai.ro, alin@racai.ro, ioana@racai.ro, sdumitrescu@racai.ro, tibi@racai.ro Abstract Spoken Language Translation is currently a hot topic in the research community. This task is very complex, involving automatic speech recognition, text-normalization and machine translation. We present our speech translation system, which was compared against the other systems participating in the IWSLT 2016 Shared Task. We introduce our ASR system for English and our MT system for English to French (En-Fr) and English to German (En-De) language pairs. Additionally, for the English to French Challenge we introduce a methodology that enables the enhancement of statistical phrase-based translation with translation equivalents deduced from monolingual corpora using neural word embedding. 1. Introduction This paper describes the RACAI entry for the IWSLT 2016 Shared Task. It consisted of three tracks: automatic speech recognition, machine translation and speech translation. The data used in the challenge came from two different domains: recorded TED talks and recorded Skype conversations1. While the recorded TED talks fall within the narrative domain, recorded Skype calls belong to the oral domain. They contain a lot of hesitations which makes it hard for speech recognition systems to successfully make use of a language model for generating correct transcriptions. Our presentation will cover an automatic speech recognition (ASR) system designed for the English TED Talks transcription track, a text- normalization technique based on a hybrid neural network/n- gram model approach and a machine translation (MT) system that convers the En-Fr and En-De language pairs. For the En- Fr language pair we took the opportunity to test a new methodology that allowed us to automatically infer and introduce new translation equivalents using data extracted from a small translation dictionary and monolingual corpora. Although we previously participated in the IWSLT Shared Task, in our previous attempts we only took part in the Machine Translation (MT) track. This is the first time we attempted other tracks in the challenge. Because of this, many of our tools and resources have not been previously evaluated on this type of data and the overall complexity of the evaluation campaign forced us to resort to baseline systems in our approach. However, this participation was important for us and provided a good contrastive evaluation metric for our future developments and participation in similar events. 2. Speech transcription For the IWSLT 2016 shared task on ASR we focused on automatically transcribing pre-recorded TED Talks. The participants were given the opportunity to experiment with 1 https://sites.google.com/site/iwsltevaluation2016/mt- track any available speech resources except for a number of TED Talks that were provided as a list by the organizers. Our speech recognition software is a standard HMM- based on the Sphinx decoder (Lamere et al., 2003). The acoustic feature vector is composed of Mel-Frequency Cepstral Coefficients (MFCC) and their delta and delta-delta coefficients. For the acoustic model we relied on the VoxForge English corpus which is composed of …

Content truncated to 500 words