Hume EVI is a voice-to-voice chat interface that analyzes the tone of your voice

[Kraut]

https://www.hume.ai/

AI Advantage:
The brilliant minds at Hume just built the most realistic AI chat experience we've ever seen. Hume EVI is a voice-to-voice chat interface, where you speak into a microphone and Hume's AI replies in its own voice in real-time.

But here's what sets it apart: Hume EVI reads and analyzes the tone of your voice!

[emk]

https://www.hume.ai/

Have you tried this? What did you think? Did you find this useful for language learning?

[Kraut]

https://www.hume.ai/

Have you tried this? What did you think? Did you find this useful for language learning?

No, not yet. Probably it's not so useful for the beginner. How can the AI react adequately to an input that is flawed grammatically and intonationally?

This is from "AI Advantage", he seems convinced:

/../
For each voice message you send, Hume tries to identify the exact emotions you're expressing, and then it uses that emotional context to help guide its responses. The result is a truly unique experience—talking with Hume felt more like a real human conversation than anything else we've experienced in the AI world.
/../

I got the following from RedBlueDebate (poe.com)

"padres helicoptero"
Blue: ¡Claro que podemos discutir en español! Pero aún así, padres helicoptero no tiene ningún sentido.
Red: ¡Oh, claro! Si quieres seguir con tus ideas locas, adelante. Pero no esperes que te tome en serio.
....

If the AI can deliver this in the "mood" of "debate", it will bring more authenticiy and emotion to the learning process.

[Cainntear]

I think this is actually kind of indicative of how there's a lot of false assumptions about AI for language learners.

The constant claim is "if it knows language, it can teach language", but very few AIs are genuinely trained to teach that. The apparent success of AI in understanding language can fall apart in cases of "hallucination", because the AI is simply identifying what it's most likely that you said. This means that it intrinsically ignores errors, and if it's ignoring errors, it doesn't know what they are. In fact AIs working to be a "partner", it will be trained to tune into your own patterns of speech, so any such AI is likely to slowly adapt into treating your characteristic errors as actually part of your personal style.

Looking at Hume EVI in that light, any early hesitancy risks establishing a baseline "normal" that will quickly become outdated as you get more fluent and confident. But then there's the issue that it probably won't handle early-stage learners well at all, possibly flagging them up as being aggressive Because They State Every Word Separately Like A Very Annoyed Person With Little Patience Might. Imagine a computer getting shirty with you and putting you off the whole experience...!

[emk]

The constant claim is "if it knows language, it can teach language", but very few AIs are genuinely trained to teach that. The apparent success of AI in understanding language can fall apart in cases of "hallucination", because the AI is simply identifying what it's most likely that you said. This means that it intrinsically ignores errors, and if it's ignoring errors, it doesn't know what they are. In fact AIs working to be a "partner", it will be trained to tune into your own patterns of speech, so any such AI is likely to slowly adapt into treating your characteristic errors as actually part of your personal style.

There is a really horrifying amount of AI snake oil out there right now. And certainly there's quite a lot for language learners. And yes, you are right that relying on AI to proofread is especially likely to produce bad results.

There are lots of useful things a language learner can do with AI. Some examples of stuff that I'm using myself:

• Reasonably accurate transcriptions of major languages for US$0.01/minute of video. This tends work well for the sort of audio people work with around B1 & B2. Once you start looking at rapid, poorly ennuciated speech and regional accents (C1 & C2 stuff, basically), it makes more errors, but it's still often better than nothing.
• Reasonably accurate subtitle translations for roughly US$0.001/minute of video.
• Grammatical explanations of what's happening in a specific sentence. With "frontier" models like GPT-4-Turbo and Claude 3 Opus, this seems to work quite well. (With GPT 3.5, you take your chances.) Again, it's not going to give 100% accuracy, but it's enough for making decent flash cards.

There are other fun use cases, like "Look at these panels of comics, transcribe the speech, and translate it."

The biggest obstacle in any AI use case is dealing with errors. It's perfectly possible to hit 95% accuracy (or even 99%) on many useful tasks. The trick is finding tasks where 95-99% accuracy is fine. For example, an audio transcription which gets 19 out of every 20 dialog lines correct is still useful. A proofreader which identifies 5% of correct sentences as errors is worse than useless.

I fully expect that somewhere from 90% to 99% of "AI" products for language learning will turn out to useless snake oil flogged to an unsuspecting public for a quick buck. When I see someone pitching a product, I ask myself:

1. What does this claim to do?
2. Does this work?
3. If this works as described, would it actually be of any benefit to language learners?

The answers to (2) and (3) are frequently "No" and "No". But there's nothing new with gimmicky methods for language learning being sold to an unsuspecting public.

[Cainntear]

There is a really horrifying amount of AI snake oil out there right now. And certainly there's quite a lot for language learners. And yes, you are right that relying on AI to proofread is especially likely to produce bad results.
...
When I see someone pitching a product, I ask myself:

1. What does this claim to do?
2. Does this work?
3. If this works as described, would it actually be of any benefit to language learners?

The answers to (2) and (3) are frequently "No" and "No". But there's nothing new with gimmicky methods for language learning being sold to an unsuspecting public.

A quicker question is "is this thing a wrapper to ChatGPT?" Normally the answer is yes, and ChatGPT can't do what they want it to, no matter how many custom instructions they add....

[bombobuffoon]

My thoughts are that I don't think "AI" can or should have any place in teaching languages, or anything for that matter.

I must admit I have wrestled with the thoughts of finding purpose for what feels so powerful and useful at first glance. All hope is washed away after a few days of use.

Now I have consigned it to the significance of a spell checker. After all what is it really, hidden beneath the facade of a chat bot? A search engine, give fuzzy questions get fuzzy answers based on datasets. Useful yes...but you will have to discover the utility for using it yourself.

[kundalini]

I came across this article on the use of ChatGPT for learning English. It isn't related to Hume EVI, but it seems relevant to the overall discussion in this thread.

https://www.donga.com/news/Society/article/all/20231110/122127200/1

“Nowadays, you can practice English conversation while driving.”

Choongin Lee (32), a graduate student in the Department of Mechanical Engineering at KAIST, has been practicing English conversation for 5 hours every day using the voice conversation function of the generative artificial intelligence (AI) ‘ChatGPT’ for a month. In a phone call with Dong-A Ilbo on the 10th, Mr. Lee said, “I took both in person and (online) video English classes, but I was dissatisfied with my progress in English.” He added, “Chat GPT can be used anytime, anywhere, and it can even teach me specialized vocabulary.” “It can be used to prepare for overseas conferences (held in English),” he said.