The idea of a human assistant for artificial intelligence that you can speak has been since the publication of “Her”, Spike Jonzes 2013 film about a man who fell in love with a Siri-like AI called Samantha. In the course of the film, the protagonist with the way Samantha, as she may appear, did not appear human and will never be human.
Twelve years later, this is no longer the stuff of science fiction. Generative AI tools such as chatt and digital assistants such as Apple's Siri and Amazon help people to maintain instructions, create grocery lists and much more. But just like Samantha, automatic speech recognition systems still cannot do everything a human listener can do.
You have probably had the frustrating experience of calling and repeating your bank or supply company so that the digital customer service bot can understand you in the other line. Perhaps you have dictated a note on your phone just to spend time to work with mutilated words.
Linguistics and computer science researchers have shown that these systems work worse for some people than for others. They tend to make more mistakes if they have a non-locals or regional accent, are black, in African-American colloquial language English, code switch, if they are a woman, are old, are too young or have a change in the speech.
Tin
In contrast to you or me, automatic speech recognition systems are not what researchers call “likeable listeners”. Instead of trying to understand them by taking other useful information such as intonation or facial gestures, they simply give up. Or you make a probabilistic guess, a movement that can sometimes lead to a mistake.
Since companies and public authorities are increasingly applying automatic speech recognition instruments to reduce costs, people have no choice than interacting with them. The more these systems are used in critical areas and range from emergency helpers to health care and law enforcement authorities, the more likely it will have serious consequences if they do not recognize what people say.
Imagine that you were injured in a car accident in the near future. You choose 911 to challenge help, but instead of being connected to a human dispatcher, you will receive a bot that is designed not to suspend any emergency calls. You need several rounds to understand, waste time and increase your fear of fear at the worst moment.
What causes this type of error? Some of the inequalities resulting from these systems are integrated into the tons of linguistic data with which developers create large -speaking models. Developers train artificial intelligence systems to understand and imitate human language by feeding large amounts of text and audio files with real human language. But whose speech do you feed them?
If a system achieved high accuracy rates in conversation with wealthy white Americans in the mid -1930s, it is reasonable to assume that it was trained with numerous audio recordings of people who fit this profile.
With strict data acquisition from a variety of sources, AI developers could reduce these errors. However, in order to build up AI systems that can understand the infinite differences in human language, which result from things such as gender, age, race, first against second language, socio -economic status, ability and much more, requires considerable resources and time.
'Right' English
For people who do not speak English – that is, most people around the world – the challenges are even greater. Most of the world's largest generative AI systems were built in English and work far better in English than in any other language. On paper, the AI has a great bourgeois translation potential and the access of people to information in different languages, but for now most languages have a smaller digital footprint, which makes it difficult for them to operate large language models.
Even within languages that are well supported by large voice models such as English and Spanish, their experience varies depending on which dialect of their language they speak.
Most speech recognition systems and generative AI chatbots currently reflect the linguistic prejudices of the data records where they are trained. They reflect prescriptive, sometimes prejudiced ideas of “correctness” in the language.
In fact, AI has demonstrated the linguistic diversity. There are now AI -Startup companies that offer to delete the accents of their users and have an impact on the assumption that their primary customers would be customer service providers with call centers abroad such as India or Philippines. The offer immortalized the idea that some accents are less valid than others.
Human connection
AI will probably be better able to process language and the processing of variables such as accents, code switching and the like. In the United States, public services under the federal law are obliged to ensure fair access to services, regardless of which language a person speaks. However, it is not clear whether this alone will be enough incentive for the Tech industry to eliminate linguistic inequalities.
Many people could prefer to speak to a real person if they ask questions about a legislative template or a medical problem or at least have the opportunity to choose interaction with automated systems when looking for key services. This does not mean that misunderstandings in interpersonal communication never occur, but when you speak to a real person, you are prepared to be a likeable listener.
At least for the moment it either works or not. If the system can process what you say, you can get started. If this is not the case, the responsibility is on you to understand yourself.
Roberto Rey Agudo, Research Assistance Professor for Spanish and Portuguese, Dartmouth College
This article will be released from the conversation under a Creative Commons license. Read the original article.
