The Linguist

20 The Linguist Vol/61 No/5 thelinguist.uberflip.com FEATURES The future of language is here, says Tanya Kendix, and the role of linguis W e know the headline: artificial intelligence is getting smarter every day, surpassing human intelligence and leaving us far behind. But before we worry about being replaced by robots, it is worth taking a look at what AI can do for us as linguists. With the rise of virtual assistants such as Alexa, Siri and Google Home, many companies are incorporating voice- and conversation-based features into their products. As this sector grows, professional linguists are increasingly involved, providing expertise and knowledge of multiple languages to deal with computational problems. Despite huge improvements in technology in recent years, the sheer complexity of speech makes it one of the most difficult things for a computer to process. Recognising spoken language and translating it into text is the role of Automatic Speech Recognition (ASR) systems – and in improving many of these systems, specialism in multiple languages is key. Accent confusion Most speech recognition systems still struggle to comprehend different accents, languages and registers. Some problematic experiences have arisen recently, including Amazon Alexa's failure to understand a Scottish accent. The issue here – despite what some sensationalist headlines may suggest – is not that robots are prejudiced. Machine learning relies on training data, in this case recordings of speech and transcriptions of those recordings. If this doesn't include enough samples of a certain accent, the AI will not acquire enough 'knowledge' of how to transcribe it, creating bias. To help limit the potential for this, linguists work alongside others during the early stages of project development. In Amazon's case, specialists in the phonetics of UK regional accents joined the team to help train Alexa to handle features of different varieties of English, such as the Scottish rolling 'r'. Professional multilingual consultants act in a similar capacity. The Google Assistant team contains experts who can improve the product's accuracy through their understanding of the nuances and accents of their languages. They also provide scope for the ASR technology to reach more languages. Specialist fields AI technologies are used in scientific and medical fields to help with tasks that humans find difficult, or that are resource intensive. The diagnosis of Developmental Language Disorder (DLD) is a prime example. Dual- language testing provides the most accurate diagnosis of DLD among bilingual people, 1 but the resources needed to support this (specialist personnel, assessment materials and software) are not always readily available. This is where an ASR system can help a clinician to make a diagnosis. A cross-disciplinary approach is required in the development of such systems, with psychologists, language specialists and programmers working together. The linguists offer their specialist knowledge of the two languages involved, especially with relation to syntax – for example by analysing the grammatical complexity of children's utterances. Such skills are also required to develop ASR technology in an educational context. For example, there is an increasing demand for research into how AI can analyse second language (L2) speech, and provide fast evaluation for more efficient and reliable grading. The development of automated assessment for L2 speakers is a growing subsector of applied linguistics. People who specialise in both of the languages being tested are able to understand the measures used in the ASR system. These consist of phonological, prosodic and temporal measures, which capture features such as pause frequency and variation in pitch and intensity. Researchers in London and Tokyo found that an ASR model could predict how listeners would judge low, mid, high and native-like comprehensibility of recorded clips of L2 speakers. 2 Transcription issues To design ASR systems like these, linguists are needed at one of the first stages: the building of the language model. Language models are used to automatically determine the next words in a given sentence. This makes it easier to identify the word being heard by comparing it to the suggestions of likely possible subsequent words. As the AI needs to be trained by transcription data from pre-existing speech recordings, professionals with specialist knowledge of different languages and dialects are needed to ensure the data used to build the models is as accurate as possible across all linguistic varieties. Transcriptions produced by professional linguists are crucial, providing all the information that the ASR system needs for training. Such transcripts have to be highly detailed and dialect specific in any language. The new frontier

• Cover