TOEFL Speaking (for the AI Era)

Automatic Assessment of Conversational Speaking Tests

My Speaking Score (TOEFL Speaking Prep) Season 1 Episode 144

Send us a text

How do AI models assess conversational speaking tests like the TOEFL and OET? In this episode, we dive deep into cutting-edge research from Cambridge University on how artificial intelligence is transforming language assessment.

We explore:
✅ How AI-powered tools like Longformer and Wave2Vec2 analyze grammar, vocabulary, pronunciation, and fluency
✅ The role of automatic speech recognition (ASR) in transcribing and evaluating speech
✅ How SpeechRater technology, developed by ETS, provides real-time feedback for test-takers
✅ The advantages and challenges of AI in language assessment and its impact on fairness and accessibility
✅ How learners can use AI tools like TOEFL Practice Online, Rosetta Stone, and Elsa Speak to improve their speaking skills

AI is revolutionizing the way we prepare for and assess speaking tests, making high-quality feedback more accessible than ever. Whether you're a TOEFL test-taker, an educator, or just curious about the intersection of AI and language learning, this episode is packed with insights to help you succeed.

🎧 Tune in now to discover how AI is shaping the future of conversational speaking tests!


Source

Free TOEFL Speaking practice:

Free resources:

Social:


My Speaking Score serves 000s of users across the globe by helping them data-power their TOEFL Speaking prep.

Hey, everybody.

(0:34 - 0:55)
Welcome back for another deep dive. Today, we're going to be talking about AI and language learning, specifically kind of those really challenging speaking tests that people have to take sometimes, like the TOEFL. We've got some super interesting research from the Cambridge University Engineering Department, and it actually focuses on health care professionals taking the Occupational English Test.

(0:55 - 1:03)
You might know that one as the OET. Yeah, that's right. But stick around, even if you aren't planning on going into medicine or anything like that.

(1:03 - 1:18)
The stuff in this research is really useful for anybody who wants to get better at speaking, especially for tests, you know, those high-stakes tests. OK, I get that. But why the OET? What's so special about that? Well, the OET, it's a lot like the speaking part of the TOEFL.

(1:18 - 1:26)
It's really focused on how you talk in conversations. It's not just about memorizing answers or anything. It's about talking with someone in real time.

(1:27 - 1:39)
And imagine you're trying to explain some complicated medical procedure in another language while still sounding natural and keeping the conversation going. That's what the OET is all about. Yeah, that makes sense.

(1:39 - 1:49)
And it's kind of like what a lot of students have to deal with on the TOEFL, right? You've got to think fast and answer questions clearly and quickly. It's not just about knowing the words. It's about actually using the language well.

(1:50 - 1:57)
Exactly. And that's where AI comes in. This research is all about how to use artificial intelligence to automatically grade these speaking tests.

(1:57 - 2:02)
And it can do way more than traditional methods. Well, wait a second. AI is grading my speaking exam? Yeah.

(2:02 - 2:09)
That's kind of scary. How does that even work? Well, it's actually a multi-stage process. It starts with what we call speaker diarization.

(2:09 - 2:18)
Basically, the AI figures out who's talking and when. So it separates what the person taking the test is saying from what the examiner is saying. Then it transcribes everything.

(2:19 - 2:26)
It turns the spoken words into text using something called automatic speech recognition, or ASR for short. OK. So far, so good.

(2:26 - 2:36)
But how does the AI actually grade the speaking? Like, that's what I really want to know. That's where things get really interesting. They use two different AI models.

(2:36 - 2:43)
The first one's called the long-former model. And it analyzes the text. It's like a super powerful grammar and vocabulary checker.

(2:43 - 2:50)
It looks for things like grammatical errors, how big your vocabulary is, and even how well you structure your answers. Wow. That's really impressive.

(2:50 - 2:58)
But just looking at the words doesn't tell you everything right. When you're speaking, things like pronunciation and how your voice goes up and down are really important, too. You're absolutely right.

(2:58 - 3:12)
And that's where the second model comes in. It's called Wave 2 Vec 2. And it works directly with the audio. So it can analyze all those subtle things that text can't capture, like pronunciation, intonation, pauses, even those little hesitations we all make.

(3:12 - 3:17)
So it's like having an AI that can hear your accent and tell you how to improve it. That's incredible. Yeah.

(3:17 - 3:35)
And what's even more amazing is that they found that combining the scores from both models, the text-based Longformer and the audio-based Wave 2 Vec 2, gave the most accurate assessment. So it's like having two expert judges, each with their own perspective, coming together to give you a really thorough evaluation. Exactly.

(3:35 - 3:42)
And this is where the connection to the TOEFL and all those challenges that students face around the world becomes really clear. OK. I'm all ears.

(3:42 - 3:57)
How does it all tie together? Well, think about all the millions of students worldwide who are preparing for tests like the TOEFL. A lot of them have limited access to good teachers or expensive tutoring services. Some might not even have reliable internet access.

(3:58 - 4:07)
AI-powered tools like this have the potential to bridge that gap. They can provide personalized feedback and support that's accessible to anyone, anywhere, anytime. Wow.

(4:07 - 4:13)
That's a game changer. So we're not just talking about AI grading exams. We're talking about AI transforming how students prepare for them.

(4:13 - 4:29)
Exactly. And the technology behind this research is already making its way into the hands of learners. Platforms that use speech reader technology that's developed by ETS, you know, the people who make the TOEFL, are using AI to give detailed feedback on spoken language.

(4:30 - 4:33)
Speech reader. I've heard of that. Isn't that used in a lot of language learning apps? Yes.

(4:33 - 4:43)
It's used in lots of different places, from apps to assessment platforms. And it's known for being accurate and for being able to pinpoint exactly what you need to improve. So this isn't just some futuristic idea.

(4:43 - 4:53)
It's something students can actually use right now. Exactly. And as this technology gets even better, we can expect to see even more sophisticated and personalized tools.

(4:54 - 5:10)
Imagine AI platforms that can analyze your practice responses, find your weak spots, and even suggest specific exercises to improve your pronunciation or grammar. This is where it gets really interesting. It's like having a virtual language coach available 24-7, guiding you every step of the way.

(5:10 - 5:19)
And for students who can't afford a tutor or don't have access to language classes, that's huge. You're right. AI is making quality language education available to everyone.

(5:20 - 5:28)
It's breaking down barriers and helping people all over the world achieve their goals. This is all super inspiring. But I'm curious to know more about the specifics of this research.

(5:28 - 5:37)
How did they actually train these AI models to be so accurate? Well, that's a great question. And it takes us deeper into the world of data science and machine learning. All right.

(5:37 - 5:51)
Let's dive in. Hashtag TTS, the Deep Dive episode 2025-0205, part 2 of 3. So to really get how these AI models work, you got to understand that they learn from tons and tons of data. In this case, they used a bunch of real OET recordings.

(5:51 - 6:01)
And they were all carefully labeled with scores given by humans. So it's like giving the AI a giant textbook of speaking examples and saying, OK, learn from this. Exactly.

(6:02 - 6:19)
They used a technique called fine tuning, which means they took a pre-trained language model like Longformer and adapted it to this specific job of scoring speaking on the OET. So like taking an already smart system and giving it specialized training to become an OET assessment expert. That's a great way to put it.

(6:19 - 6:33)
And for the audio model Wave2Vec2, they used something called self-supervised learning. It's pretty amazing, actually. The model learns general representations of speech sounds and patterns from a massive data set of audio without any direct human input.

(6:33 - 6:48)
So the AI is basically teaching itself to hear and understand spoken language, kind of like a baby learning to talk by listening to everyone around them. Exactly. And then just like with Longformer, they fine-tuned Wave2Vec2 using that OET audio data.

(6:48 - 7:01)
So it learned to focus on the acoustic features that match up with high speaking scores. Like sending the AI to speech school, first learning the basics, and then getting specialized training to be an OET scoring pro. That's a fun way to think about it.

(7:01 - 7:14)
Now remember, all this was done with the OET, which focuses specifically on health care communication. But imagine what this could mean for a test like the TOEFL, where the language use is even more diverse. That's what blows my mind.

(7:14 - 7:34)
If AI can be this accurate with the OET, think of what it could do for TOEFL test takers all over the world. It opens up so many possibilities, especially for people who don't have access to traditional learning resources. Imagine AI platforms that can give you real-time feedback on your pronunciation, grammar, even vocabulary, all tailored to your specific needs.

(7:35 - 7:40)
Like having a personal tutor available 24-7. Exactly. And it's not just about individual learners either.

(7:40 - 8:02)
AI could also be used to make large-scale language assessments like the TOEFL, more efficient and fair. Think about automated scoring systems that could eliminate human bias and give more consistent scores across different testing centers and examiners. So we're talking about a potential paradigm shift in how we approach language learning and assessment, making it more accessible, personalized, and ultimately more effective.

(8:03 - 8:09)
Precisely. And this leads us to the broader impact of AI on education as a whole. It's not just about language learning.

(8:09 - 8:22)
AI has the potential to revolutionize how we teach and learn across all subjects. I can definitely see that. Personalized learning paths, adaptive assessments, even AI tutors that can adjust their teaching style to each student.

(8:23 - 8:39)
The possibilities are endless. And the most exciting part is this technology is developing so quickly, we're only just beginning to see what AI can do for education. So this OET research is really just a glimpse into the future of AI-powered language learning, a future where language barriers are broken down.

(8:39 - 8:48)
Access to quality education is available to everyone. And learners are empowered to reach their full potential. It's a future we're really excited to be a part of.

(8:48 - 9:04)
Me too. But before we get too carried away with all the possibilities, let's bring it back down to earth for a minute. For our listeners who are studying for exams like the TOEFL, what are some practical things they can take away from this research? What can they do right now to use the power of AI? That's a great question.

(9:05 - 9:15)
And it's one we'll explore in more detail in the final part of our deep dive. Hashtag TTS, the deep dive episode 2025, Zo2-over-5, part three of three. Welcome back to our deep dive.

(9:16 - 9:24)
You know, we've been exploring this fascinating world of AI and language learning. We've looked at some groundbreaking research. We've gone into the details of these AI models.

(9:24 - 9:32)
And now it's time to get practical. Yeah, you're right. Let's take all this knowledge and turn it into something our listeners can actually use, especially those who are preparing for tests like the TOEFL.

(9:32 - 9:57)
So for someone who's studying for the TOEFL right now, where should they start? How can they actually use AI to get better at speaking? Well, the good news is that AI-powered language learning tools are becoming more and more accessible. There are so many options out there, from apps to online platforms that use AI. Yeah, but it can be a little overwhelming, right? Like, where do you even begin? Absolutely.

(9:57 - 10:11)
I'd say a good place to start is to look for platforms that use speech rater technology. Remember, that's the engine developed by ETS, the people who make the TOEFL. Right, so it's like tapping into the same expertise that's used to actually score the test.

(10:11 - 10:26)
Exactly. And platforms that use speech rater often give you really detailed feedback on different parts of your speaking, like your pronunciation, fluency, grammar, and even vocabulary. So you're not just getting a score, you're getting specific insights into what you're doing well and what you need to work on.

(10:26 - 10:35)
That's it. It's like having a personal coach who can pinpoint exactly what you need to focus on. Some platforms even give you targeted exercises based on how you speak.

(10:35 - 10:46)
Now, that's what I call personalized learning. Are there any specific examples you can give us, just so our listeners have a better idea of what's out there? Sure. A popular one is TOEFL Practice Online, or TPO.

(10:46 - 10:59)
It's offered by ETS themselves, and it lets you take realist and practice tests and get feedback from speech rater. That sounds super helpful, especially for getting used to the format of the test and the types of questions you might get. It is.

(10:59 - 11:12)
Another interesting example is Rosetta Stone. They've added speech rater to their language learning programs, so you get feedback on your pronunciation, it helps you fine-tune your accent and speak more clearly. Rosetta Stone, that's a classic.

(11:12 - 11:26)
I remember seeing their ads when I was a kid. It's amazing how they've evolved to include this cutting-edge AI technology. It really shows how AI is changing the world of language learning, and there are many other platforms out there too, like Duolingo Babble and Elsa Speak.

(11:26 - 11:38)
They're all using AI in different ways. It's like a whole smorgasbord of AI-powered options. So with all these tools available, it sounds like there's no reason why anyone can't achieve their language learning goals, right? Well, almost.

(11:39 - 11:51)
Remember, AI is a powerful tool, but it's not magic. It can't replace the hard work and dedication that it takes to really learn a language. Right, you still need to put in the time and effort to practice, learn new words, and immerse yourself in the language.

(11:52 - 12:00)
Exactly. Think of AI as your personal language learning assistant. It's there to guide you, support you, and give you feedback along the way.

(12:00 - 12:15)
It helps you streamline the process, make it more efficient, and even more fun. So to all our listeners out there, especially those who are working hard to master the TOEFL Speaking section, embrace the power of AI. It's an incredibly valuable resource on your language learning journey.

(12:16 - 12:24)
Absolutely. And remember, the journey is just as important as the destination. Enjoy the process of learning, embrace the challenges, and celebrate every step of the way.

(12:24 - 12:33)
Well said. That wraps up our deep dive into the world of AI and language learning. We hope you found it insightful, inspiring, and most importantly, empowering.

(12:34 - 12:41)
Until next time, keep those language learning goals in mind, stay curious, and keep exploring the amazing world of languages. Happy learning!

People on this episode