Just a word of caution regarding text-to-speech programs (Neospeech, etc.).

I thought it would be an awesome idea to start memorizing near native sounding sentences using text-to-speech (TTS) software. I proceeded to do so. The result, however, was disaster, because about 85% of the time the voices are very natural, the other 15% of the time the voices are unnatural, for example, inflecting up when infection should be down, inflecting down when inflection should be up, unnatural pacing, etc. And since I had memorized the sentences, my pronunciation was messed up.

I had a hard time unlearning the pronunciation of the TTS software (which I thought sounded excellent), but eventually I did. How? My Japanese teacher spent 30 minutes a week with me practicing pronunciation and even volunteered to record sentences, pleading with me to stop using the TTS voices. He said when they were on they were really on, but when they were off they were terrible, and "やめた方がいい"
. I was sad because I thought I had found a nice, cheap solution to making sample sentences to practice, but it turned out too good to be true.

Here are the voices I have, all of which are according to my teacher not up to snuff for using to practice Japanese:

Neospeech Show, Miyu (purchased from nextup.com), Misaki. Misaki is the voice that is probably around 85% correct, Miyu and Show have their strengths/weaknesses but are probably more like 50-60% correct.

I see buried in the AJATT monster thread http://forum.koohii.com/showthread.php?tid=682&page=18 a similar discussion, but figured this topic warrants its own space. And it looks like some native Japanese people give Misaki the nod, but if achieving natural Japanese is in your set of goals, you probably would be better off finding examples sentences to practice elsewhere.

Maybe its not so good for pronunciation, but 85 percent might be high enough to help with listening comprehension without causing too many problems. However, reading your post definitely makes me want to reconsider my learning plan as far as how much priority to put on TTS. Maybe I should try to find Japanese audio books instead and then follow along with the printed versions. I would consider ripping audio from my favorite DVD movies, but I would want transcripts to go with them, and that would probably be a lot of work. Being able to SRS audio seems to be such a benefit; its too bad there aren't more sources.

Dragg Wrote:Being able to SRS audio seems to be such a benefit; its too bad there aren't more sources.

だから一緒に協力してがんばりましょう！

I find the whole idea of using TTS for language learning crazy. These things always sound bad and it's much better to learn from text only than from a bad robotic voice, at least you don't reinforce bad habits. I'm really surprised you even tried it.

Codexus Wrote:I'm really surprised you even tried it.

Yeah, the reason why I tried it is because the English voices to me sound really good. I thought to myself, "If my Japanese friends could pronounce English as well as these voices, I'd be amazed, and the Japanese voices must be equally as good..."

I had actually used the English voices to study some stuff in English (on the go) so I figured they might be worth using - the ATT natural voices actually sound really natural. Although, now that I think about it, I often used the voice editor to make the pronunciation of the English voices more natural since out of the box didn't work as well as I wanted them to. But, to be fair, a lot of what I was studying had medical vocabulary in it - none of the voices could say "clinical" right, or E. coli...

Hi, I'm glad I read this thread. I was planning on using Japanese TTS as part of my study plan, but now I'm having second thoughts. I had the same impression as you Jay, I was blown away by how good a lot of the english voices sounded (especially the different accents), and thought the japanese voices would be equally as good. If anyone can give me some information on the following points that would be helpful.
1. I was hoping to use the TTS to make audio files so I could study on the go. Can the TTS programs on the market take a file that has say 1000 rows and 2 columns (1 Japanese sentence, 1 english sentence), and then produce an mp3 file or a series of files which contains the information? ie Can you tell the software: process this file, and use english for this field, japanese for that field etc.
2. How does it know what reading to use for kanji?
3. If the english is so good, how could the makers stuff up the japanese. Especially since the pronunciation is relatively regular?
4. I wonder if it is better to learn with TTS and then correct any errors in your pronounciation later, or is it better not to use it at all?

It is too bad there ain't more audio materials available for learning japanese. To get good at listening and pronouncing you need good audio materials to practice with. The CDs that come with textbooks are often too limited. Random dramas / animes etc are too unstructured for learning, especially early in your study. Stuff like Pimsleur is too basic.

I have to admit I stopped using the TTS a while back for a few reasons, but the major one was that sometimes it does make mistakes on inflection, simply because the front end program gets readings wrong, and passes those incorrect readings on to the TTS engine.

If it doesn't read the kanji you feed it properly, and just goes off of kana, it usually blows the pronunciation, from my experiences.

I'd say it's a neat toy to play with, and could be very useful as a news article reader for when you're working out, but the front end needs to be improved so it doesn't screw up readings.

I'm still trying to find a copy of that NHK pronunciation dictionary on CD-ROM. Dang thing is out of print. >_<

Codexus Wrote:I find the whole idea of using TTS for language learning crazy. These things always sound bad and it's much better to learn from text only than from a bad robotic voice, at least you don't reinforce bad habits. I'm really surprised you even tried it.

I don't agree with the idea of modeling yourself on tts either, but Misaki is VERY lifelike. Enough to fool even natives much of the time. Show on the otherhand is a joke.

I use Misaki on the answer side of my sentence questions. I don't have to read the kana to check if I read all the kanji correctly but I just listen to the voice.
I know there are mistakes like wrong kanji readings but it doesn't really matter for me since I can always check the sentence written in kana. I think it's a gread addition as I can practice listening comprehension like that. Even if the mistakes (like wrong inflection) make my Japanese sound unnatural, I think it is still better than trying to pronounce everything on my own. Like this I can at least get 85% correct.
Surely TTS can't replace real Japanese audio input but you can still watch dramas and anime or listen to Japanese podcasts or Japanesepod101. I will continue using TTS and can recommend it to anyone else. It makes reviewing my sentences much more fun.

I don't advise learning from Misaki, but even though I already know a decent amount of Japanese they are still words I come across that I find hard to pronounce 輸出（ゆしゅつ）comes to mind. But using Misaki to and mimicking her I can now pronounce it properly thankfully

If you do decide to use tts as a study tool, I can recommend nextup.com and their text aloud program - very easy to convert things to mp3.

And I disagree that memorizing sentences isn't helpful. I think it is extremely helpful - both for allowing native-like production and improving listening ability - I just wish I had time to do it for every word I come across!

I agree that Misaki is great 85% of the time. When she's "on", she's indistinguishable from a native speaker. (I've tested this with a native speaker.) The other 15% of the time, she either selects the wrong reading or has bad inflection. When she selects the wrong reading, the problem can be partially corrected by replacing the kanji with kana, but of course that sometimes creates an inflection problem as she can't identify word boundaries. I've listened to enough real Japanese that I can usually tell when Misaki is messing up, so I think the damage to my listening and/or pronounciation skills is limited.

To me, the benefit of having comprehensible audio to match the sentences in my SRS outweighs any problems that Misaki might introduce. I've created listening flashcards for all of my reading flaschards, and the listening cards seem to be helping my listening comprehension. Having TTS cards also adds some variety to my SRS reps.

MeNoSavvy Wrote:1. I was hoping to use the TTS to make audio files so I could study on the go. Can the TTS programs on the market take a file that has say 1000 rows and 2 columns (1 Japanese sentence, 1 english sentence), and then produce an mp3 file or a series of files which contains the information? ie Can you tell the software: process this file, and use english for this field, japanese for that field etc.

The TextAloud software many people use with Misaki does have some batch processing features. It will split a large file into separate files and then batch-process those individual files into mp3s. I "retrofitted" hundreds of sentences that were already in my SRS this way, though a little knowledge of SQL and scripting was necessary to get the sentences in and out of Anki and to associate the newly created audio files with their respective flashcards.

MeNoSavvy Wrote:2. How does it know what reading to use for kanji?

I don't know exactly how it knows. I seems to make an educated guess. It's usually right but is sometimes wrong. When it's wrong, you can replace the kanji with kana and re-record the sentence, but that can create create other problems ("correct" but odd-sounding pronounciation).

MeNoSavvy Wrote:3. If the english is so good, how could the makers stuff up the japanese. Especially since the pronunciation is relatively regular?

I doubt that the Japanese is any worse than the English. In my opinion, Japanese is probably easier to pronounce than English, though I'm sure each language has idiosyncrasies that make realistic-sounding TTS a challenge.

MeNoSavvy Wrote:4. I wonder if it is better to learn with TTS and then correct any errors in your pronounciation later, or is it better not to use it at all?

This is a matter of opinion. I think learning with TTS helps my listening comprehension and spices up my reviews, but others have different and equally valid opinions to the contrary. I'd suggest trying it out and deciding for yourself.

Codexus Wrote:I find the whole idea of using TTS for language learning crazy. These things always sound bad and it's much better to learn from text only than from a bad robotic voice, at least you don't reinforce bad habits. I'm really surprised you even tried it.

If you haven't heard the best TTS engines recently, then I can see why you think of us as crazy. But they have come a long way since the Stephen Hawking robot-voices of the 90's. There are moments when, at least to my ear, they sound exactly like real people.
If the trend continues, I won't be surprised if virtually perfect voices come out within the next 5 to 10 years.

I have also tried and enjoyed the AT&T Natural Voices, and if the Misaki voice is comparable in terms of accuracy, then I would probably still recommend such programs to learners, although I am on the fence still. Basically, if ever, it should probably just make up a small portion of a well balanced study program. Additionally, it would need to be phased out eventually in favor of listening to 100 percent native Japanese.

Jarvik7 Wrote:I don't agree with the idea of modeling yourself on tts either, but Misaki is VERY lifelike. Enough to fool even natives much of the time.

Where do you get this information from? I tried the website and I'm not convinced. It's good but not that good.

So far I prefer using audio rips from anime/drama/movies. It's much more safe pronunciation-wise and also more fun if you like those shows. It's also not very difficult to do with the right programs.
Of course the downside is that you don't get the audio for every sentence you can think of, but tv shows usually contain a hell of a lot of sentences anyway!

Here are the results of a little test on a Japanese friend, asking 1) how does Misaki sound and 2) what is she saying?

Quote:Sounds like computer sound. Sometimes it's hard to hear because of the noise.
しにいたる、きょくたん、しゃど、しゃどのあたりが発音がちょっとおかしい。
しゃど、のところは何を言ってるか良くわからない。

深刻　アメリカ　しにいたる　死にそう　患者　病気　根底　極端　橋　箸　雨　飴　若い　和解　しゃど　斜度　娘　医療事故　理解　事件　歩いていた

What Misaki was actually pronouncing:
深刻　アメリカ　死に至る　死にそう　患者　病気　根底　極端　橋　箸　雨　飴　若い　和解　車道　斜度　娘　医療事故　理解　事件　歩いている。

So in this little test, she manages natural pronunciation 80% of the time, with unnatural or difficult to understand pronunciation the other 20%. I was impressed she could get 若い　and 和解 correct.

Noise? Do you mean noise in the listening environment or coming from the TTS? I hear no noise here... While there are certainly times where it sounds slightly computery, giving it a list of words to speak is not a very realistic test. You should have given it sentences to speak instead. Should also be a blind test instead of "does this computer sound like a computer?".

For me, I don't think I'm trying to learn to speak by listening to the computer. I'm just using it to give me audible dictation. It's just a different input channel, aural vs. visual. I'm not really trying to find a TTS voice that is indistinguishable from a human speaker (though naturally that wouldn't hurt). Y'all seem to be searching for a very high standard of quality that really isn't necessary for my purposes. Good enough is good enough.

To revive an old thread, I have seen many people talk about Text-to-Speech programs in studying. Past the beginner stage, is using them in studying beneficial in your opinion? To those few that have heard me (I know Tobberoth will reply), do you think it would help or hurt? Even if I just listen to it, will it eventually cause me to sound more unnatural or will it boost my listening comprehension?

You might find 'proper' shadowing to be more beneficial:
http://forum.koohii.com/showthread.php?p...2#pid46892

Also Text to Speech can help you remember your kanji stories:
http://forum.koohii.com/showthread.php?p...2#pid45462

As a heads up, the latest post on AJATT talks about TTS for Cantonese (and Japanese to a lesser extent).

How do you do when the TTS doesn't read the kanji correctly? For example 方 by itself is different when talking about direction or a person. Do you convert those words to kana when the TTS screws up?

Khatzu Wrote:...Not only that, but they have no “learning” ability - you can’t “teach” (customize) them to pronounce certain things correctly. Misaki does have such ability; she can even be “taught” intonation…

Khatzu says that we can teach Misaki to pronunce certain things correctly, and that she can be taught intonation also. How can we do this?

Like Ahibba, I too would love to hear more about how to program intonation into the Misaki voice. Also, where did you guys track down this "Misaki" voice anyways? It's not for sale on the nextup website....

To all the people that integrate TTS with their studies- What TTS software do you use and where did you purchase/find it?