Microsoft Speech Language Translation (MSLT) Corpus: The IWSLT 2016 release for English, French and German

Microsoft Speech Language Translation (MSLT) Corpus: The IWSLT 2016 release for English, French and German Christian Federmann, William D. Lewis Microsoft Translator Microsoft Research, Redmond, WA, USA {chrife,wilewis}@microsoft.com Abstract We describe the Microsoft Speech Language Translation (MSLT) corpus, which was created in order to evaluate end- to-end conversational speech translation quality. The corpus was created from actual conversations over Skype, and we provide details on the recording setup and the different lay- ers of associated text data. The corpus release includes Test and Dev sets with reference transcripts for speech recogni- tion. Additionally, cleaned up transcripts and reference trans- lations are available for evaluation of machine translation quality. The IWSLT 2016 release described here includes the source audio, raw transcripts, cleaned up transcripts, and translations to or from English for both French and German. 1. Introduction We describe the Microsoft Speech Language Translation (MSLT) corpus that we created to evaluate end-to-end qual- ity of Microsoft Translator’s speech translation system, the system powering speech translation in Skype Translator. The goal of Skype Translator is to support open-domain, sponta- neous conversations between individuals who speak different languages, ultimately such that one would not be able to tell the difference (e.g., semantically or in latency) between con- versations held in one language and those held in two. We carefully constructed the corpus so that the Test and Dev sets would not be constrained by the current state-of-the-art; in effect, we wanted our test data to represent the gold standard of our ultimate aspirations for speech language translation. The MSLT corpus currently contains full end-to-end speech translation Test and Dev sets for three languages: English, French and German. To adequately test the con- versational translation scenario, and any components used in speech translation, e.g., Automatic Speech Recognition (ASR), disfluency processing, Machine Translation (MT), etc., whether combined or in isolation, the MSLT corpus consists of the following data: audio files, their verbatim transcriptions (i.e., faithfully capturing the speech input, including disfluencies), cleaned-up transcripts (i.e., where disfluencies are removed), and translations. The corpus contains subcorpora for each pair in and out of English (so, English→French, French→English, English→German, German→English). This paper presents details on the recording, transcription and translation processes we con- ducted to build this corpus. 2. Data Collection 2.1. Recording Modes For our experiments with speech translation we attempted to create realistic Test and Dev sets. Crucially, we wanted test data that represented actual bilingual conversations, and was not constrained by the current state-of-the-art or do- main.1 Ultimately, it is our research aspiration to develop open-domain speech translation that differs little with respect to domain, latency, etc., from monolingual conversations or from bilingual conversations between fluent bilinguals. Thus, we wanted test data that came as close as possible to fully bilingual conversations, or fluently translated monolin- gual conversations. Initially, we paired consultants and tested several recording scenarios in a limited study, comparing dif- ferent modes of conversations between two speakers: 1. Monolingual conversations between English speakers 2. Bilingual conversations between English and Spanish speakers with …

Content truncated to 500 words