https://www.duolingo.com/profile/AnCatDubh

Everything the pronunciation gets wrong in the very first lesson (also what’s with the audio?)

  • Short vowels should be −ATR:
    • /a/ [a] — correct
    • /e/ [ε] — inconsistent
    • /i/ [ɪ] — off
    • /o/ [ɔ] — off
    • /u/ [ʊ] — blatantly off
  • Long vowels (other than /aː/) should be +ATR:
    • /eː/ [eː] — sounds a bit inaccurate
    • /iː/ [iː] — correct
    • /oː/ [oː] — off
    • /uː/ [uː] — haven’t heard yet, I think
  • When a vowel appears before a nasal consonant, it should be nasalized, and if the consonant is intonation phrase-final or followed by a fricative, it should be omitted with the preceding vowel lengthened — haven’t heard yet (should listen for how sum ‘am’ is pronounced)
  • /l/ should be pronounced [ɫ] unless before /e(ː)/ or /i(ː)/ — off
  • Latin stress rule: the accent falls on the penultimate (i.e. second-last) syllable if it’s heavy, or the antepenultimate (i.e. third-last) syllable if the penultimate one is light — inconsistent
  • /s/ was retracted, i.e. [s̺] — off

So we get words like:

Also I’m pretty sure Latin sentences did not have a consistently obvious echo effect or end with a loud clicking sound of a recording device getting switched off.

This is really lazy work. It’s somehow even worse than the Ukrainian audio and the (old) Irish audio.

Also, some of the sentences have missing audio because of course they do.

Get it together, Duolingo. You’ve been releasing half-baked courses over and over recently and it’s getting ridiculous.

August 31, 2019

32 Comments


https://www.duolingo.com/profile/Anbrutal

A double /ll/ is regularly [l:], likely a palatalised [lʲ:] - so [puˈεl.la] is right and [pʊˈεɫ.ɫa] is wrong. The only possible exception might be when it results from assimilation, e.g. cum lūnā might just have been [kʊ̃ɫ.ɫu:na:] for those speakers whose lūna started with a velar [ɫ] (whose number was reducing with time judging by the changing testimonies and the Romance developments). It's very unlikely, but not entirely impossible. Also, prevocalic vowels are regularly tense, so [pu], not [pʊ].

In general your critique is well-deserved and well-informed, but it misses the elephant in the room - vowel length is completely disregarded. None of the speakers seemed to me like they had the vaguest idea that it even exists. While vowel qualities make or (especially for English speakers) break your accent, the absence of vowel length makes and breaks the entire language that is Latin. Mixing them up is precisely equivalent to mixing up every short vowel /e/ with every dipthong /ae/ - and if you were going to say "but the Romance languages got rid of vowel length" - yes, and this was tightly connected with the loss of almost the entire case system, because vowel length played a central role in distinguishing the forms. They also mixed every short vowel /e/ with every dipthong /ae/. At that point, the language was no longer Latin.

Nobody can claim to be using the Restituted pronunciation scheme while pronouncing half of the vowels the same as the other half.

August 31, 2019

https://www.duolingo.com/profile/JimKillock

For me this problem of vowel length goes hand in hand with leaving out vowel-length markers (macrons). I can't see how a learner is meant to understand and learn that In Italiā habitō is pronounced with a long a if it is not marked, and inconsistently said. There's more on this tho, which I will post in a separate thread.

I thought I could hear vowel length differences in the male voice at least some of the time btw. I will keep listening though to see if this is the case.

August 31, 2019

https://www.duolingo.com/profile/Anbrutal

The male speaker who has a distinctly Italian intonation only seems to reproduce vowel length because he has a tendency to lengthen final vowels, as well stressed vowels in words of more than 2 syllables. More often than not this happens to be correct or at least not as wrong as the other way around... but I can assure you he doesn't distinguish short and long phonemes.

August 31, 2019

https://www.duolingo.com/profile/JimKillock

It would be a lot easier to tell whether there are errors if the text had macrons, as I'm having to 'hear' and sometimes I won't really know … I am sure that having macrons would also help the audio recorders to self-correct.

August 31, 2019

https://www.duolingo.com/profile/Anbrutal

Oh, no doubt. Just like it would be difficult to spot and correct missing double consonants if for some reason the course creators decided "You know what? Let's not spell those." Which the Romans themselves didn't like to do, so the precedent is there, just like with vowel length. Macrons absolutely need to be any kind of learning material, not least here on Duolingo.

August 31, 2019

https://www.duolingo.com/profile/Sonnenkrieger

In general your critique is well-deserved and well-informed, but it misses the elephant in the room - vowel length is completely disregarded.

I completed about 1/3 of the course just to try it out. This very fact blows my mind. I don't know how this audio passed any sort of inspection by knowledgable Latin speakers. Yes, this is beta, but that is a tired excuse. What has to happen now is that basically all of the sentences have to be re-recorded because they didn't get it right the first time.

August 31, 2019

https://www.duolingo.com/profile/JimKillock

As I go on the audio does seem to get better; there are I think two female voices, one is considerably better than the other (or else, perhaps, the same person eliminated most of the US intonations). The male voice does distinguish vowel length in my estimation.

August 31, 2019

https://www.duolingo.com/profile/ClarkJaman

Yes. I won't be learning Latin on DL until they fix the terrible audio.

September 12, 2019

https://www.duolingo.com/profile/Semeltin

I completely agree with you! Well, except that I think the assimilation you mentioned isn't that unlikely. :P It comes down to if the /m/ was still present in the underlying representation.

Two relevant discussions:
https://forum.duolingo.com/comment/33881462

https://forum.duolingo.com/comment/33618227

September 1, 2019

https://www.duolingo.com/profile/Hermesianax

and if you were going to say "but the Romance languages got rid of vowel length" - yes, and this was tightly connected with the loss of almost the entire case system, because vowel length played a central role in distinguishing the forms.

Loss of vowel length played a role in the drop of the case system, sure. However, it's really only the long and short -a in the first declension that depends on vowel length. Just look at Greek, which also lost vowel length, but retained three of its four cases. "Central role" is a bit of a hyperbole. I think the nasalisation and eventual loss of -m played a bigger role in this development.

August 31, 2019

https://www.duolingo.com/profile/Anbrutal

It's not just inside a single paradigm, and admittedly what I said is more valid in a system where the elimination of the contrast results in an ē/i, ō/u and ae/e merger (which is what happens in the vast majority of Romance languages), so facilis and facilēs becomes /faceles/, magnus and magnōs /magnos/ (and if you cut the /s/ in the singular, you run into homophony with the ablative magnō), sōle and sōlae become /solε/ and pretty quickly /sole/ (not to mention what happens after cutting the /s/).

As a result, the same ending is shared by adverbs from two different declensions, 3d decl ablatives, 1st decl genitives, datives and plural nominatives, and imperatives from two different conjugations, and combined with the open-syllable lengthening you can't tell whether someone is telling you "morning!" or "stay!". Even supposing a Sardinian-type merger will net you sentences like "The butt gave a gift" instead of "(S)he gave the children a gift" (Donum natis dedit) - pretty sure Medieval Latin is full of puns like that.

You can find a nice list of Sardinian-type minimal pairs in the manual called Prosōdia Latīna, and one could easily collect such a list for general Romance as well (obviously excluding all the ēs/is and ōs/us, which are basically infinite).

As for Greek, the fact that it lost vowel length yet retained the cases nicely illustrates the point I'm making about the Latin declension system. And as for nasalisation, actually it manifested itself rather like modern phonosyntactic doubling in Italian - so wasn't length per se, but still had direct relation to syllable weight. This is why it's often not represented in writing from the earliest times.

September 1, 2019

https://www.duolingo.com/profile/Hermesianax

You make excellent points on vowel quality being responsible for a lot of merging of case endings, but that's still not the same as vowel quantity. I don't see how that changes my initial point.

As for Greek, the fact that it lost vowel length yet retained the cases nicely illustrates the point I'm making about the Latin declension system.

In what way? I think the example of Greek shows that it isn't primarily quantity loss that causes case endings to merge.

By the way, may I ask what you mean by "cutting the final -s"? Because if we look at Spanish and French, it's clear that the drop of final -s was not a universal characteristic of sound development in Vulgar Latin. Again, nasalisation of the -m and eventual drop was more significant because the accusative had become the casus generale in late antiquity.

September 2, 2019

https://www.duolingo.com/profile/ARCANA-MVSA

It's a serious hyperbole. There are other overlapping cases where the pronunciation was the same (dative and ablative plurals, looking at you - and first declension nominative plural, genitive singular, and dative singular too). There are some words where the macron distinguishes between tense (venio, venire, veni, ventus is one of these), but these are not only uncommon, but it's also a piece of cake to tell from the context which one it is without the macron. (Context is, arguably, Duolingo's biggest weakness.)

As you said, the only case that truly relies on the macron is first declension ablative; but really, even then, in my experience, it's not too hard to figure out if it's nominative or ablative without the macron. Distinguishing between the three -ae forms is actually harder.

Volgav vitsenanieff nivya kevach varatsach.

August 31, 2019

https://www.duolingo.com/profile/Semeltin

"/a/ [a] — correct" – Did you forget about /aː/? /maː.tɛr/ : /pa.tɛr/.

I wouldn't criticise them as much if the audio reflected the correct phonemes. That's what's really important because you have to learn the correct phonemes for every single word. And as Anbrutal pointed out, you can be wrong about an allophone even if you are well informed about the phonology.

September 1, 2019

https://www.duolingo.com/profile/Leandro_lhi

Isn't the audio often made by a program? Why bother with people recording it if it's gonna sound unprofessional? Might this change after the Beta? I still appreciate all the hard work and it's good to have it in whatever form it is, I just wonder about this decision

August 31, 2019

https://www.duolingo.com/profile/Robert-Alexan

Your are all way to harsh. This is a beta. Manage your expectations. (And yes, I also don't like the often too American way of pronouncing many words).

And understand that this issue is a linguistic minefield. Although scholars agree on the pronunciation in the Classical period for the most part, you have to acknowledge that there is still some issues that spark heated debates. And be realistic: even the most brilliant scholar or voice actor can't hide his accent. That's just too much to ask. I will always be able to say where the speaker is from... approximately. Unfortunately Romans didn't upload podcasts to Spotify back then.

There's no good TTS for Latin. Best in town is the mbrola voice la-1 that can be added to TTS software ESpeak. It tries to honor vocal length and stress, but ain't perfect. It's robotic, fails for some letter combinations and doesn't do the nasals (although the extent of nasals is subject to discussion). It also fails for Greek words with the letter combination ph (which were presumably pronounced the Greek way by erudite Romans) and didn't get nihil and mihi right (although scholars can argue about the right pronunciation for the later anyway). And it needs macrons to work correctly - "macronizing" a text is time consuming.

In general, you also produce decent results when using a modified text with an Italian TTS. But that requires a lot of work, since you have to rewrite the whole Latin text in order to approximate Classical pronunciation. And results are not always predictable and you have to try all kinds of weird stuff like throwing in some hyphens, h here and there or doubling some letters in order to force the right stress. And the end result will always be a bit too much on the ecclesiastical side of the trench.

August 31, 2019

https://www.duolingo.com/profile/ARCANA-MVSA

You spared me the effort of trying not to get too annoyed while explaining the same thing. Thank you.

Some people clearly have no idea what "Beta" means. Or handmade audio, for that matter.

P.S. Your Spotify comment made me chuckle. :)

Volgav vitsenanieff nivya kevach varatsach.

August 31, 2019

https://www.duolingo.com/profile/Danielconcasco

Thank you! To echo what Arcana-mvsa said, you saved me a lot of typing.

August 31, 2019

https://www.duolingo.com/profile/Robert-Alexan

(Shameless self-promotion following). Check out my attempt on Latin TTS:

https://www.memrise.com/course/1480193/5000-most-frequent-latin-words-audio/

In most cases (especially in the 2nd half of the course) there's exactly two recordings for each word: the Mbrola robot voice and an attempt to trick various Italian TTS systems into speaking Classical Latin. Both are acceptable for the most part. Check out the syllabus for sources, approach and limitations. I experimented with dozens of different speech synthesizers and different sources. Needless to say that results were a mixed bag. Even tried to record some words myself - nah, too Germanic. I know my limits. "Macronizing" the list was painstaking. But hey, it's a 5000 most frequent word list with acceptable audio, macrons and extended info on gender, POS, declension and conjugation. That's not nothing, I suppose.

September 1, 2019

https://www.duolingo.com/profile/JimKillock

Do you have sufficient permissions to release the list publicly in some way? This seems like a great resource.

That is, as a public resource / database, eg via Pastebin, Github or similar. I can think of some useful ways to deploy this info, for instance as a check to find missing words in Wiktionary to help prioritise new entries.

September 1, 2019

https://www.duolingo.com/profile/JimKillock

Hi Robert: probably. The EU has "database right" which can get in the way. In the USA there's no restriction on reuse of data which are mere facts.

As most of your sources are out of copyright this may be ok even in the EU. However if I extracted your whole database without your permission I'd certainly be breaching EU DB right.

September 1, 2019

https://www.duolingo.com/profile/JimKillock

Thanks very much Robert!

(As an aside, any kind of curated, structured data qualifies as a database; arguably there is database right in a card index file, or even in a well organised shelf of books.)

September 1, 2019

https://www.duolingo.com/profile/Robert-Alexan

It's public ain't it?

Copyright issues involve tricky questions. Plz see my sources and judge for yourself, if that's good enough for you. I think, it's OK if you state your sources, didn't just copy a single other database with no own adjustments whatsoever, and treat this as an academic work compiled on the basis of various other freely available academic papers for the sole purpose of teaching and learning- at least in EU copyright law (I'm not a legal expert though). I'm OK with any form of reuse. There's add-ons for Memrise that extract all entries incl. audio.

September 1, 2019

https://www.duolingo.com/profile/Robert-Alexan

@Jim

[...] EU [...] database right [...]

Not so sure if my vocab qualifies as a database. However, that's why I've said I'm okay with any reuse of my course if you state your sources.

September 1, 2019

https://www.duolingo.com/profile/ARCANA-MVSA

Latin, along with the Irish course, does not have a TTS voice (the computer-generated voice you're probably thinking of). This results in handmade audio clips that will certainly contain their share of errors and "echo sound effects" (actually the result of speaking into a microphone) and probably occasionally missing audio, as the contributors may have missed a spot or two, or may have removed one and are awaiting a replacement. (Or it might be like the Irish course, in which many of the sentences are missing audio because they simply couldn't get to them all.)

Robert-Alexan explained quite well why Latin doesn't have a TTS voice and also why it may or may not get one in the future. But I think it's also worth noting that no matter the quality, it's still surprising that Latin has audio at all - most courses are initially released without any. That in itself is impressive.

As for the accent ... quite frankly, best-case scenario would be to have a native Latin speaker make the clips. Unfortunately, there hasn't been one of those around for the past however many hundreds of years. Which means, naturally, that all we can do is approximate and try to imitate what we think the language would have sounded like - which leaves ample room for one's own accent to peer through, since of course we've got no inkling of what their accent would have sounded like.

It's not that it sounds unprofessional, it's imperfect. Just like it always will be to someone or other. But I still at least appreciate the effort.

Volgav vitsenanieff nivya kevach varatsach.

August 31, 2019

https://www.duolingo.com/profile/JimKillock

I don't really agree: there are a lot of very good Latin speakers who are very attentive to getting it right. It's not impossible at all. FWIW it is the female voice which has most errors while the male voice is pretty good.

For the most part, teachers are not inclined to get their accents correct, as it is such a low priority for courses. Thus we have a lot of people with not so good Latin accents around, who are otherwise incredibly talented and knowledgeable.

This is exacerbated when you can 'hear' the speakers' native accent, whether English, American English or something else. Unfortunately placing these inaccuracies into a global context like a Duolingo course is going to expose them very quickly.

The lazy way to do it is to ask Italian or Spanish native speakers to help with the audio (maybe the male voice here has [some] Italian for instance?). They'll have less problem with the vowels, which makes it sound better almost instantly, although they can struggle with using 'w' and Italians have a different 'r' as I understand it.

As an additional point leaving out macrons is a bad idea because it makes learning the sounds much harder. With them it would also be easy to see where the speakers have made mistakes. In the medium term macrons help a lot with distinguishing some otherwise confusing apparent duplications, including in tenses. It's a shame these aren't present.

Wiktionary has a good collection of recordings for many common Latin words. You can hear some of these here: https://en.wikiversity.org/wiki/Latin/Basics_Lesson_1

https://en.wikiversity.org/wiki/Latin/Basics_Lesson_2

Those btw are the lessons from CarpeLanam, when she helped out with a Duo-style Latin course, which covers most of the grammar now, excepting the subjunctive. Sometime I hope these lessons will contain audio for the sentences, which may help people following the new Duo course while it is still quite short.

August 31, 2019

https://www.duolingo.com/profile/ARCANA-MVSA

"Getting it right" can mean very different things to different people. That is my point.

Another part of the problem is that there's a lot of people here complaining about the quality, but I don't see them volunteering to contribute ...

Volgav vitsenanieff nivya kevach varatsach.

August 31, 2019

https://www.duolingo.com/profile/Danielconcasco

I'm sure they already sent sound files samples of their pronunciation to the staff along with a schedule of when they are available to record ;)

August 31, 2019

https://www.duolingo.com/profile/JimKillock

I don't think there is an option to do that. I am not sure if my own pronunciation is up to it; but when I am confident that it is, I will certainly be contributing Latin audio somewhere, although it is more likely to be through Wikimedia Commons to Wiktionary and Wikiversity, as these are well set up for collaborative volunteer contributions (they do also have editorial control to remove things that are wrong).

As an aside I did port CarpeLanam's Latin lessons across from these forums to Wikiversity – with her kind permission – and then added the existing audio from Wikimedia Commons to it, along with internal links to Wiktionary and macrons for about a third of the course, so I think it is fair to say I've done quite a bit on this front.

August 31, 2019

https://www.duolingo.com/profile/AnCatDubh

I say get bilingual English–Lombard speakers.

August 31, 2019

https://www.duolingo.com/profile/AaronD.2

While I will contend that the Calabrese system is the best for vowel pronunciation, I will happily give you all 400 of my lingots for pointing out the disgrace of ignoring vowel length.

August 31, 2019

https://www.duolingo.com/profile/Kati673185

speakers are really bad -

August 31, 2019
Learn Latin in just 5 minutes a day. For free.