The Linguist is a languages magazine for professional linguists, translators, interpreters, language professionals, language teachers, trainers, students and academics with articles on translation, interpreting, business, government, technology
Issue link: https://thelinguist.uberflip.com/i/1229313
add, substitute and replace words. The transcript needs to be accurate enough so that someone using the edited transcript for translation purposes will not need to consult the audio file. Task 2: ASR + machine translation The translator was given an audio file as well as the corrected ASR output from task 1. Task 2 consists of fully correcting the ASR output after it has already undergone basic level correction. Capitalisation, agreement and accents need to b e corrected. Sentence boundaries need to be inserted for systems such as Google ASR and corrected for others, including IBM ASR. This stage consists of a complete and thorough transcript correction so that it is presentable to an MT engine. The corrected text is submitted to Google Translate and the translator then works on the MT output, measuring the time it takes to post-edit it. This is a standard workflow in machine translation and is referred to as post-editing machine translation (PEMT). Task 3: Speech-to-text translation This scenario consists of using an end-to-end speech-to-text translation system, such as Enrique Peña Nieto 2 and Serbian President Aleksandar Vučić. 3 We also used a TV interview with the Bolivian Economics Minister Luis Arce 4 in order to cover dialect variation. We designed three tasks representing different workflows, measuring the time it took translators to achieve correct translations in each scenario and comparing the results to find out which was quickest. The experiments were performed by four translators: one French, two Spanish and one Serbian. For each task, the audio file was run through IBM ASR and Google ASR, except for Serbian, which was processed only by Google. Google Translate provided two outputs: the ASR transcription output and the translation output, which we are calling 'speech translation'. Task 1: ASR + human translation The translator was given an audio file which had been run through the ASR system. This task consists of 1) correcting the ASR output and 2) translating the document into English. Note that correcting the ASR output for this task is time-consuming. It involves using a media player, listening to the audio, comparing it with the ASR output, and going forwards and backwards with the player to 22 The Linguist Vol/59 No/2 2020 ciol.org.uk FEATURES Can a combination of machine translation and speech recognition the audio translation process? Evelyne Tzoukermann and Corey A udio transcription and translation are at least 2-4 times as time-consuming as text translation. Finding ways to make the process more efficient, therefore, would be of great help to translators working in this field. The time it takes to translate or transcribe audio is vastly dependent on audio quality, which can be impaired by background noise, overlapping voices and other acoustic conditions. So could the process be speeded up using a combination of automatic speech recognition (ASR) and machine translation (MT)? We set out to evaluate how useful these technologies can be, using time and quality of translation as the primary factors. We chose to work with French, Spanish and Serbian using two publicly available online systems: IBM Bluemix and Google Translate. There are still well-known issues with these services, despite dramatic improvements, but they offer sufficient capability for the purposes of our study. For French, we selected excerpts from French President Emmanuel Macron's 2017 inauguration speech, 1 segmenting the file into two parts of 60 and 120 seconds. For Spanish and Serbian, we selected similar political speeches by former Mexican President RACE AGAINST THE MACHINE © SHUTTERSTOCK