mariano54

Just added this to my benchmark site: https://multilingualsttbench.com/

It doesn't reach the frontier in either latency or accuracy for ai multilingual conversations.

I'm sorry but we are not providing longer free trials at the moment. You can clear the cache to try again (but you will have to go through the onboarding).

You can do this in the web app: https://app.issen.com/manage-account

Thanks for the feedback.

1. The response latency / patience can be configured in the quick settings (tap the gear during the call).

2. There already is a bias in transcription. It's something we're actively improving, multilingual transcription isn't a solved problem yet

3. It is supposed to correct you actively if you have corrections enabled (check the settings). But we received feedback that it's not enough, so we will ramp it up soon

Yes it pushes the problem, but it's a much easier problem, and models like Gemini flash 2.5 do very well.

The mobile app can be used in a car (not the Web app). Looking at the screen is not required.

Minimax's new model is quite good. We use their voices for some of our Japanese tutors. The pitch accent is almost perfect.

There are incorrect reading or Chinese readings occasionally, but you can tell when that happens due to the furigana being different

I'm sorry about that. May I ask what language you tried?

Thanks for pointing that out. Did you try the other ones by any chance?

Langua is a nice app with more features than ISSEN. We focus on the voice chats, and I think our tech is better (intelligence, latency, realtime flow, voices, etc), especially for non english/spanish.

We are aiming to create a long term companion/tutor that gets to know you more and more, and can create customized curriculums, lessons, etc.

Our market is different - ISSEN is more focused on intermediate and advanced students, and especially those that are willing to use a more difficult app in order to learn faster.

Also, having the AI voice tutor as our main feature allows us to iterate quickly, and be well positioned for future improvements in AI models.

As for marketing and GTM, we're in the super early stages, and there is definitely a lot of competition out there, it won't be easy.

The persuasion lesson sounds like a great idea, we haven't thought of that. Yeah voice to voice models will be amazing. There is significant progress from openai/gemini, and we plan to use them when they are ready.

Exactly, it's all server side. There are no plans for this. The main issue I see with doing it in device is the LLM piece. Even with some large models like llama 4 maverick, the tutor just struggles to properly teach and understand the student, it's not viable IMO.

Intelligence is super key here, especially as the context size gets larger (due to memory) and intelligence degrades.

Another major issue is TTS voice quality, but this seems to be improving a lot for small local models.

EDIT: You're right, latency is also a big deal. You need to get each piece under a second, and the LLM part would be especially slow on mobile devices.

Thank you so much for that!

We focused on testing and tweaking the most popular ones, we have not tested some of the niche ones. We have removed languages that users have told us have major issues, but there are still some left.

The voices are due to the quality of the TTS services that we use. Openi, 11labs, minimax. Some services don't have many or even 1 good voice. We will add more over time

Sesame also passes in the users voice into the TTS model so that it can vibe well with the users tone and mood, whereas we are just using raw TTS. Their latency is also very low, but this is not quite suitable for language learning.

In the future we hope to move to full voice to voice models, once those become mature and intelligent enough.

The issue with that is people with less common names will get the wrong transcription, and there's nothing worse that seeing your name spelled wrong over and over

May I ask which tutor you were using? The TTSs have different strengths and weaknesses.

Good idea on the export, will add that to our to-do list.

Ouch it looks like all the personality promoting has made it confused about its identity.. thanks for pointing it out

Thanks! No, it cannot do this yet, as it's using a pipeline of voice to text to voice. I think models are heading in that direction, and voice to voice models are getting better.

For pitch accent, shadowing is a great way to improve. You can pause and repeat the tutors messages for example, or read out the word when doing flashcard reviews (copying the flashcard audio).

It's less tolerant of mispronunciations. There is custom promting to explicitly leave in mistakes and to not fix them. It's still not perfect and it (the speech to text module) sometimes corrects the user's pronunciation mistakes.

Yes we do have this issue, but it's improved a bit over chatgpt due to using multiple transcribers.

The models are improving though, and they are at a very good place for English at the moment. I expect by next year we will switch over to full voice to voice models.

I think so. Language learning is not just about practicality, it's about knowledge, culture, respect, and connection between people of different countries. If people have more free time and travel more, I expect language learning to grow.

Plus, you can't do auto translation for languages like Japanese where the grammar is reversed. Auto translation has fundamental limitations.

Thanks for the feedback. You might want to try manual mode, where you press enter or say "send" to trigger a response

You should try re-generating a custom curriculum from the settings, and there you can prompt it to create specific lesson types.

We are really sorry for the subpar experience. We did not test Vietnamese, and it seems like the quality is not sufficient. Thanks for pointing out the issues.

Yeah, agreed, we started with a similar observation. These voice models are getting better quickly.

You do need an app to create a holistic learning experience just for language learning. Customized curriculum, tons of prompting, AI models chosen for transcription accuracy, flashcards/dictionary, etc.

We also support hands free mode, and many other things are customizable like slang, speaking speed, target language usage, etc.

There is a structured curriculum that gets generated after the intro lesson (if you responded yes to the curriculum question).

This is available for all proficiencies. It's just much harder to talk for hours in a new language as a beginner. It's usable but requires more effort.

Some random tips which could help you

- Don't go over 10k tokens in the prompts as the intelligence and memory degrades

- Summarize sessions and save the summaries, potentially summarize the summaries as well

- use VAPI or realtime api if you want to build fast. Building the full pipeline takes a while

- try out different models and see how personality varies. Our favorite is gpt4.1 with temperature 1.

- goal system. The promot should always contain the current goal, and the next goal. Evaluate goals with another LLM, and dynamically change the prompt

Yeah this is a really great idea, perhaps you can sample tokens based on some kind of SRS and I+1 sentence system. Kind of like a graded reader but for speaking. Will definitely look into this in the future

ISSEN is designed from the ground up for this use case.

* curriculum, completely customizable, with grammar, roleplay, topics, speaking speech, transcript, dictionary, corrections, etc

* prompting and AI models all chosen to be a better fit for multilingual, easy to understand, etc.

* the tutor actively tries to teach you, it's not an assistant

* integrated flashcards that go hand in hand with the speaking immersion

That's correct.