AI Pronunciation Practice
Something is not working? Check the trouble shooting section below.
***
(The recording is limited to 20 seconds (for now) and the character limit is 200. That’s enough for a fairly long sentence. But don’t copy paste in large chunks of text ;).
Viel Spaß and leave a comment how you like it.
Options
Edit sentence
***
YourDailyGerman Pronunciation Practice – FAQ
What kind of AI is that?
The AI is a deep neural net that was developed and trained by a small team at EF Languages specifically as a tool to get feedback on ones pronunciation. It takes a written German sentence as input and it grades your rhythm, speed, melody and of course the quality of sounds. It does so for every single word and also gives an overall rating.
Note that it’s NOT based on one single recording. There are many ways to say something and get 100% score.
How good is the AI?
It does a surprisingly good job at grading native speakers, even if they mumble or swallow sounds. The AI still grades them near 100% because it has learned that this is how native speakers sound.
However, it has some weak spots here and there. For instance, I have trouble with the word “du”, which I have to pronounce overly clearly to get a good rating.
But the system generally gives you a good indication of how well you sound.
What do the colors mean?
The AI grades in percentages (x% of native speaker), and I have assigned the colors based on those percentages. The cutoff between the levels is somewhat arbitrary. Here’s an overview.
For the words:
Below the words, you can see a breakdown of the individual sounds and how clean they were. Hovering/tapping on them will show you the percentage. The word grading are an average of the sounds, so if you get all sounds perfect except one, the overall word score will be less than perfect.
What is “good”?
If you can get it to green, that means you’re somewhere between 80 and 99 percent. That is good enough. Of course you can try to go for 100%, but for some words it probably won’t work.
I am a native speaker and I get 100% most of the time, but not always.
How do I know what the problem is?
That’s a bit beyond what the tool can give you at the moment. The tool is a guide to which sounds you need to work on, but if you really want feedback on how exactly you’re doing them wrong and how to improve them, you need to work with an actual person.
Trouble Shooting
I can’t record anything!
The browser needs permission to access the microphone. It might be disabled for this website. To enable it, click on the lock icon next to the URL. There, you can enable it.
My mic is enabled and working, but it’s still not working
Try a different browser. If that is not working, chances are that you are using an iOS device (iPhone, iPad). Apple is not able and willing to provide the standard mic support classes that every other system uses. It works on some websites and doesn’t on others and it tends to not work here.
I can’t really do much about it, so maybe just try another device. The issue has been known since 2017 at least and Apple doesn’t give a sh*.
I get very poor results
If you’re getting poor results, chances are that the recording quality is not very good. Play back some recordings and listen to how they sound. If they’re muffled or noisy, that has a huge influence on the grading. Try to get closer to the mic if possible.
***
If you have any questions, feedback or ideas please leave a comment. Always happy to upgrade the feature.
Hi! My AI gives example pronunciation with a very heavy American accent, how can I change that and get proper german examples?
This is weird, I tried multiple times just now and it sounds fine for me. It shouldn’t be based on what your mothertongue is or anything like that.
Do you hear a male or a female voice?
I hear a male voice
Can’t play the audios on my Macbook and Android phone :/
I think I just fixed it. Could you try again, but possibly in an incognito browser?
Yes, its working now! Thanks a lot!
The short “I” and long “I” seem to show up as being pronounced correctly even if I say them the exact same way. Specifically, if I use the long ‘I” for both. For example, the “ie” in “Mieten” can be used to pronounce the “I” in “Ich” without any problems. I would still get 100% percent, despite those clearly being different sounds. Does this work for you.
I’m not asking if it is correct to pronounce them the same in Standard german. (It isn’t), but does this work when you try it using the pronunciation checker. And do you know why this is happening?
My guess would be that they didn’t train that properly. It seems to be consistent so the AI really doesn’t have a concept of the short and dark “i”.
Same for bitte vs biete.
Interesting!!
So it does notice when the prompt is “mieten” and I say “mitten” instead. It marks the “ie” and the singular score was like 50%.
But when I tried prompt “mitten” and said “mieten”, it does reliably give 100% on the “i” which is BAD.
That’s the first real glitch for this AI. They’d probably have to adjust and train that a bit.
Thanks for the response. Yeah, that makes sense. Hopefully it get‘s fixed.
Like most things AI, it works only in theory. I can’t get more than 0% no matter what I do. I don’t say Icm the best speaker in the world, but I have C1, I live in Germany and people understand me
This is likely not an AI issue but a recording issue.
Did you try to playback your recording? What do you hear?
My guess is that you’re on Apple with iOS. Audio recording doesn’t work for some of them because they don’t support the quasi standard that everyone else is using.
I figured it out. I’m using Firefox. I had to click on the shield icon up left by the address bar and turn “enhanced tracking protection” off
Ah sweet!
I’ll keep that in mind if someone else has an issue with Firefox in the future. Glad you got it to work!
I just cannot get the ü sound. Any ideas how I can form this sound.
So generally, you make a “uuuu” sound, round your lips a lot for it. And then you move your tongue forward and up, like you want to squeeze a chewing gum at the top of you mouth behind the teeth with the tongue. Doing that will change the u to ü.
On individual words (by themselves). Would natives generally get 100% fluency and accuracy?
Actually, let me rephrase. The fluency meter in general is confusing. It is significantly easier to achieve a higher fluency score when saying a full sentence than it is when you just say a single word. It you type in a single word, and then repeat a sentence with that same word in it. The word might show up with a 50% on fluency. But then when you type the full sentence in. The full sentence might have a 100% fluency.
I think fluency for a single word doesn’t really make sense as the tool uses pauses between words as one dimension.
Just ignore that score for single words please, it doesn’t mean much.
Makes sense
Just tried it and no, I only got 60%.
Quick question. Does a 100% actually mean the pronunciation is like a native? I feel like I get 100% on individual words too frequently. And sometimes I swear I didn‘t say it in a way that sounded native. Does 100% just mean your pronunciation is “good enough”?
100% is definitely very very good, it’s not just “good enough”, but more like “like a native”.
The model basically says that it matches 100% with what it “expects” from a native speaker.
If you have a way to share a recording with me, I’m happy to give it a listen and give you feedback.
Well, the easiest example would be that I can just say “nine” in English (I am a native English speaker), and I will get 100% on “Nein”.
Yeah, it definitely is too forgiving when it comes to pronunciation. I can just speak English, and it will give some perfect sounds.
(as in I can just speak English, and certain sounds will get 100%. I can say “No” and get a 100% on “N”. This shouldn’t happen, as the German “N’ should be more dental than the English “N”. This happens with a lot of other sounds.
When I type in the letter “e” and nothing else. I always get 50% accurate and 0% fluent. No matter how I make the sound, if I totally say something different, or literal make no sound at all. But it works fine with other letters like “i” or “o”.
Same for me.
It’s a neural net AI so it may not do well with individual letters. A letter is not a constant, but it’s ending transforms toward the sound that follows it.
Maybe the AI picks up on that and because it’s missing for a single “e” it doesn’t get “realistic” vibes.
It’s not meant for single letters. Try “es” instead and “See” for the two qualities.
Thanks for the answer. Makes sense. I‘ll try it
is it possible to have the german examples in male as well as female voices?
It’s definitely possible but I will have to switch away from 11labs first. It’s too expensive with the current volume what people produce here and in the text comprehension tool :)
If there was one thing my German teachers hammered into me over the years, it was the uvular (or “guttural”) r, particularly at the front of words. When I listen to your pronunciation, as well as the folks on Easy German, I can identify it clearly. You may not be able to answer this since you can’t hear me, but the “r” pronuncation by the LLM (eg “rot”, or “Rosen”) sound much less emphatic , almost like “wot” or “Wosen”. Are you able to comment at all? Thank you so much. I really love this thing.
To me, the LLM audio “r” sounds normal, like a native speaker, I have to say.
Thanks!
What does it mean when a word is gray? Like accuracy wise? Unless it doesn’t have anything to do with accuracy?
Nvm. I should have looked earlier. I didn’t notice you updated the page.
Yeah, I changed the “adequate” color from green to gray because there was a bit too much green going on. I’m thinking about changing the percatage cutoffs too, but I’m not sure what the best way is.
Hi Emanuel
Great update!
Using a Pixel 8pro, repeating the same sentence:
Phone hands free, 10cm to inbuilt mic ≥ 80%
Yealink BH71 Headset ≥ 50%
Google Pixel buds A series ≥ 70%
I encourage users to try memorise the sentence and not simple read it – to get a better Sprachgefühl. Ya gotta mean it, embody it.
“Ya gotta mean it, embody it.”
Love it, this is the way!!
You don’t want to be a person who reads the sentence off of a learning website. You want to be a person who SAYS the sentence in real life :).
Thanks for the comparison of devices. I found that the laptop mic is awful on my computer whereas the phone mic works great and gives me great scores.
Seriously though I think these interactive elements are a bonus for learners and thanks for integrating them!
Hi Emanuel,
The more I use this the more I like it! I do have one question. Below the initial evaluation of what I’ve spoken (with the various shades of green and occasional yellow) is a second version; same sentence of course but the letters are gray with occasional yellow in apparently random yellows. I’m not sure what that second version is supposed to represent. I might have two very high scores, but that second line is still there so it doesn’t seem to be related to the scores.
I added this just a few days ago and haven’t written a description yet.
That’s basically the sound for sound grading of your speech.
The sounds get the same grading that the words do, but I have decided to only use two colors. Anything above 70% match is gray and the yellow ones are the “problem sounds” in that word. So if you get less than perfect score in a word, the highlighted sound are what you should focus on. OR check for them in the recording. It may always be the recording that’s cuasing the grade, so don’t be too harsh on yourself.
thanks! mystery solved!
Based on the colors of the words, “accuracy” seems to be the score you’re referring to. But what does “fluency” measure?
Fluency measures how many or few breaks you make. You can try it out by making one longer pause somewhere. Speed doesn’t matter as a much as a continued stream of speech.
It all worked for me and I am delighted to have a way to improve speaking ability! This new toy makes me ridiculously happy. Beverly. Rochester, NY
Hi Emanuel The layout overlaps in Firefox so I have changed to Chrome and the layout is fine. Whilst most words are green, several letters are red or yellow and, after multiple tries, there seems no consistency on which letters go red and yellow. Should I be worried about this or is this where I should focus my pronunciation to get all the individual letters green? Regards Peter
Might be a css caching issue on firefox that resolves itself. Or it’s Firefox being “extra” again :).
The letter display just shows you which letters/sounds the AI wasn’t sure about. The word ranking is an average of all the letters. So it might be green, but one letter is yellow.
I’m rolling out an update later today that makes the letter display less aggressive.
Note that this CAN also be due to recording quality. If I record on my computer built in mic, I do get yellow letters sometimes. Use them as orientation, but it’s not 100% reliable.
Hi Enanuel. I think there must be ESP between us as I ask if this was coming only a couple of weeks ago and you have delivered! It is a good confidence booster to see that most words are green. However, some of the individual letters are red and yellow so I can see which ones I need to work on. Thanks and regards, Peter, Brisbane.
I think our talking about it brought it back into focus and I realized how much better it would be. Originally, I was just gonna pull examples from my example database, but then I had this idea with focus on sounds. It’s a much much better tool now, so thanks for being the catalyst.
The random sentence generation works, but I cannot enter my own sentences. I can add a sentence in the editor, but it does not register it when I hit enter. Am I missing something?
The way it is right now is that whatever is in that text-box is sent to the AI along with your recording. If you enter your own sentence, that’s just going to live in the textbox for now. I have not added audio creation for that.
Just enter something and hit record and you’ll be graded based on what you entered.
Ich habe es! es funktioniert. Cool.
I really admire your courage, trying to get something this complex to work on many devices!
Brilliant idea.
“Other”: I speak or type sentences into Google Translate on my phone. Yeah, no sound.
I can‘t get it to give me a grade on either my iPhone or iPad. I have iOs Version 26.3. I researched online how to rectify, like turning off pop-up blockers, etc. but nothing works. It‘s so frustrating. I did try it on my desktop and it works perfectly. Unfortunately, my desktop is not in a private area where I can practice. I hope Apple fixes this issue.
They will not fix this, they do not care.
Could you tell me what exactly you’re seeing?
1) Does the generation work (text and audio)
2) DOes the recording button change color when pressed?
3) Does a sound clip of you pop up (a play button)
4) If yes, do you hear something?
5) What else happens on screen when you start and stop recording?
I can’t make promises but this might help.
1. Yes. It generated a sentence and played it well.
2. The recording button on the iPad does not change color when pressed the same way it does on my desktop. On the iPad it flashes green but on the desktop it‘s a solid red.
3. Yes a sound clip pops up but it says next to it: Accuracy: NaN% and Fluency 0%
4. When I hit play next to my sound clip I do not hear anything.
5. Nothing else happens on the screen.
Yeah, that’s the classic Apple problem then. I’ll look into it again, but I can’t make any promises because I cannot test and debug without an iPhone.