This is a great initiative and I hope to see more come out of this; I am not criticizing, but just want to provide my user experience here so you have data points.

In short, my experience lines up with your native speakers.

I found that it loses track of the phonemes when speaking quickly, and tones don't seem to line up when speaking at normal conversational speed.

For example, if I say 他是我的朋友 at normal conversational speed, it will assign `de` to 我, sometimes it interprets that I didn't have the retroflexive in `shi` and renders it `si`. Listened back to make sure I said everything, the phonemes are there in the recording, but the UI displays the wrong phonemes and tones.

By contrast, if I speak slowly and really push each tone, the phonemes and tones all register correctly.

Also, is this taking into account tone transformation? Example, third tones (bottom out tone) tend to smoosh into a second tone (rising) when multiple third tones are spoken in a row. Sometimes the first tone influences the next tone slightly, etc.

Again, great initiative, but I think it needs a way to deal with speech that is conversationally spoken and maybe even slurred a bit due to the nature of conversational level speech.

There's one thing that gave me pause: In the phrase 我想学中文 it identified "wén" as "guó". While my pronunciation isn't perfect, there's no way that what I said is closer to "guó" than to "wén".

This indicates to me that the model learned word structures instead of tones here. "Zhōng guó" probably appears in the training data a lot, so the model has a bias towards recognizing that.

- Edit -

From the blog post:

> If my tone is wrong, I don’t want the model to guess what I meant. I want it to tell me what I actually said.

Your architecture also doesn't tell you what you actually said. It just maps what you said to the likeliest of the 1254 syllables that you allow. For example, it couldn't tell you that you said "wi" or "wr" instead of "wo", because those syllables don't exist in your setup.

Though, as a guy who speaks perfect mandarin from Beijing, I’m struggle even to pass the easy ones… So it can definitely used some improvements. The example 你好吃饭了吗 returns hào → hǎo, fān → fàn, le → liǎo. The first two are the model listen my tone mistakenly, and the last one should be le instead of liǎo in this context.

Also I see in the comment section people are worry about tones. I can guarantee tones are not particularly useful and you can communicate with native speakers with all the tones messed up and that’s perfectly fine. Because as soon as you leave Beijing, you’ll find all the tones are shuffled because of every region has their own dialect and accents, which doesn’t stop people from communicate at all. So don’t let tone stuff slow your learning process down.

It helped a lot even if I did look like an insane expat conducting an invisible orchestra.

One more thing: there's quite a bit of variation in how regional accents in the mainland can affect tonal pronunciation. It might be worth reaching to some native speakers to give you some baseline figures.

If you can’t easily hear your pronunciation mistakes so clearly it hurts, consider putting more energy into training your ear. Adult language learners usually have brains that have become resistant to, but not incapable of, changing the parts of the brain responsible for phoneme recognition. The neuroplasticity is still there but it needs some nudging with focused exercises that make it clear to your brain exactly what the problem is. Minimal pair recognition drills, for example, are a great place to start.

It’s not the most fun task, but it’s worth it. You will tighten the pronunciation practice feedback loop much more than is possible with external feedback, so a better accent is the most obvious benefit. But beyond that, it will make a night and day difference for your listening comprehension. And that will get you access to more interesting learning materials sooner. Which hopefully increases your enjoyment and hence your time on task. Plus, more accurate and automatic phoneme recognition leaves more neurological resources free for processing other aspects of your input materials. So it may even help speed things like vocabulary and grammar acquisition.

( I’m learning using a flashcards web app I made and continue to update with vocab I encounter or need: https://memalign.github.io/m/mandarin/cards/index.html )

The thing you've built is so good, and I would have loved to have it when I was learning Mandarin.

I tried it with a couple of sentences and it did a good job of identifying which tones were off.

I am mostly developing this for myself, to have the perfect tool for me, but I dare say, that I have not seen anything comparable and that I let my 10y+ experience in learning Chinese influence my design decisions. Oh, and it is free /libre software of course (AGPL). It comes with an ever improving vocabulary file that has tons of metadata about words, their usage, how to memorize them, etc. under ODbL (open database license).

[1]: https://codeberg.org/ZelphirKaltstahl/xiaolong-dictionary

Pronunciation correction is an insanely underdeveloped field. Hit me up via email/twitter/discord (my bio) if you're interested in collabing.

[0]: https://gist.github.com/anchpop/acbfb6599ce8c273cc89c7d1bb36...

Just curious - would you need insane HW infrastructure to begin with, or hosted/managed. And what tooling is preferred by the industry for the "training"?

There are still holdouts!

Come back to me in a couple of decades when the trove of humanity's data has been pored over and drifted further out of sync with (verifiable) reality.

Hand-tuning is the only way to make progress when you've hit a domain's limits. Go deep and have fun.

I'm sure there are a bunch of apps out there that claim they do the same thing, but they don't, IMO. Even if they do, as you said, where is the fun in that?

Great post, thanks for it!

I played around with python scripts for the same purpose. The AI gives feedback that can be transformed to a percentage of correctness. One annoyance is that for Mandarin, the percentage is calculated at the character level, whereas with English, it gives you a more granular score at the phoneme level.

I suck at chinese but I want to get better and I'm too embarassed to try and talk with real people and practise.

This is a great compromise. even just practising for a few minutes I already feel way more confident based on its feedback, and I feel like I know more about the details of pronunciation.

I'm worried this might get too big and start sucking like everything else.

https://github.com/sequoia-hope/mandarin-practice

For me it doesn't work very well. Even easy phrases like 他很忙 get transcribed completely random "ma he yu". Is it maybe over-fitted to some type of voice?

It would be cool if a model could tell you if you are singing or playing a piece of music with the right intonation and other ways.

- māmā (incorrect)

- māma (correct)