#Can you confirm if lexicons/aliases are
meager pike
They should be applied from what I see in the code.
on-the-fly generation from a Project WebUI chapter - does it mean pressing Play/Regenerate by selecting a paragraph, or pressing Download chapter?
Would you be able to prepare minimalistic example?
One paragraph, dictionary with one rule, result and expected result, so I can test it?
Do you see same issue using Play/Regenerate and Download chapter (on freshly created paragraph)?
Sorry that you are running into this issue, my suspicion is that it's our numbers normalizer that got introduced recently and it's working well for the top used languages. Maybe it doesn't know when to use Swiss-French, I'll look into adding option to turn it off. But in the meantime it should be possible to instead of converting paragraph by paragraph, pressing Download chapter when you don't have paragraph converted - I don't think this workflow uses new numbers normalizer yet.
abstract forge
I mean regenarating the paragraph in a chapter of the project in web ui.
I can prepare a test gladly. How should I share the project with you ? Just paste the url in the location bar with the project open ?
Thank you for your answer btw !
meager pike
If you can copy the pronunciation dictionary and text example that should be enough.
abstract forge
I've started with the test, but I see that in context ("the shoes are seventy francs") or "bare" ("seventy") the results are very different...
I'll maybe add more "chapters" in the project, with the different approaches. But we definitely had in-context numbers that were pronounced as the wrong number, in short sentences though. A table broken into chapter where the whole chapter would just be eg "Main account : seventy-two point two francs" (translated for you).
I'll be interrupted for about an hour now. Will get back to you after that
Ah, do you mean you just need the text, not the project reference ?
meager pike
Let me rephrase, I think both could work, either project url that is generating wrong audio, or just copy paste text here (in french) and pronunciation dictionary that you are using, so I can test it in my own account. But if it's easier Project url should work too.
abstract forge
Ok, I've tried 2 types of text, "bare numbers" and "in context number", and in three "modes" : without lexicon, with a phonetical alias lexicon, with a "correct form" alias lexicon (just repeating the same word and spelling).
Here are the texts :
"Bare numbers" :
Et voilà , en français (intro pour donner le contexte de la langue) :
septante
septante et un
septante-deux
septante-trois
septante-quatre
septante-cinq
septante-six
septante-sept
septante-huit
septante-neuf
"Numbers in context" :
Et voilà , en français (intro pour donner le contexte de la langue) :
Le nombre de participants est de septante.
On est Ă septante et un kilomĂštres de Lausanne.
AprĂšs septante-deux minutes, le fichier Ă©tait prĂȘt.
Le menu coûte septante-trois francs.
Les septante-quatre personnes invitées sont toutes là .
Si les septante-cinq passagers sont lĂ , nous sommes prĂȘts.
A septante-six, cela commencerait Ă faire beaucoup de monde.
DĂšs septante-sept, cela serait trop.
S'il y en a plus de septante-huit, nous pouvons faire deux voyages.
Et voici le dernier chiffre : septante-neuf.
The "phoenetical aliases" lexicon :
<?xml version="1.0" encoding="UTF-8"?>
<lexicon version="1.0"
xmlns="http://www.w3.org/2005/01/pronunciation-lexicon"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://www.w3.org/2005/01/pronunciation-lexicon
http://www.w3.org/TR/2007/CR-pronunciation-lexicon-20071212/pls.xsd"
alphabet="ipa" xml:lang="en-GB">
<lexeme>
<grapheme>septante et un</grapheme>
<alias>septantéin</alias>
<grapheme>septante-deux</grapheme>
<alias>septantedeu</alias>
<grapheme>septante-trois</grapheme>
<alias>septantetroi</alias>
<grapheme>septante-quatre</grapheme>
<alias>septantequatre</alias>
<grapheme>septante-cinq</grapheme>
<alias>septantecinq</alias>
<grapheme>septante-six</grapheme>
<alias>septantesisse</alias>
<grapheme>septante-sept</grapheme>
<alias>septantesette</alias>
<grapheme>septante-huit</grapheme>
<alias>septanteuitte</alias>
<grapheme>septante-neuf</grapheme>
<alias>septanteneuf</alias>
</lexeme>
</lexicon>
The "correct form" lexicon :
<?xml version="1.0" encoding="UTF-8"?>
<lexicon version="1.0"
xmlns="http://www.w3.org/2005/01/pronunciation-lexicon"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://www.w3.org/2005/01/pronunciation-lexicon
http://www.w3.org/TR/2007/CR-pronunciation-lexicon-20071212/pls.xsd"
alphabet="ipa" xml:lang="en-GB">
<lexeme>
<grapheme>septante et un</grapheme>
<alias>septante et un</alias>
<grapheme>septante-deux</grapheme>
<alias>septante-deux</alias>
<grapheme>septante-trois</grapheme>
<alias>septante-trois</alias>
<grapheme>septante-quatre</grapheme>
<alias>septante-quatre</alias>
<grapheme>septante-cinq</grapheme>
<alias>septante-cinq</alias>
<grapheme>septante-six</grapheme>
<alias>septante-six</alias>
<grapheme>septante-sept</grapheme>
<alias>septante-sept</alias>
<grapheme>septante-huit</grapheme>
<alias>septante-huit</alias>
<grapheme>septante-neuf</grapheme>
<alias>septante-neuf</alias>
</lexeme>
</lexicon>
And here is a table of the results. You'll see, it varies wildly....
NB and this is just for the 70-79 range. The 80-99 range has similar results that I didn't test in that systematic way.
abstract forge
Oh, I've just realized you mentionned this :
"But in the meantime it should be possible to instead of converting paragraph by paragraph, pressing Download chapter when you don't have paragraph converted - I don't think this workflow uses new numbers normalizer yet."
The first test I did worked, I hope I can finish the job with that trick đ€
We'd be very happy with a switch for that setting ! Other less known languages too I guess.
abstract forge
But it very unpractical, because now I have to find a way to "reset" every chapter of a 30+ chapter project, and as far as I know, there's not other way to do this but cut and repaste the text of every chapter đ
For this, and other needs, a search function would be sooo handy !
meager pike
I'll look into adding setting to disable it in paragraph as well later this week!
abstract forge
Thank you so much !
abstract forge
I've clicked on "Download project zip" yesterday night, got the "it can take a few minutes" message and I went home.
This morning I got the same message. I thought that maybe the audio was indeed generated, but now the zip still needed to be generated. But after 30min I was still getting the same message.
I asked your discord bot for hints, and it mentionned disableing the volume normalization. I did it and I was now able to download the project.
I've checked the audio, and the issue with the number seems to have been resolved (number normalized skipped) with this approach. Thanks for the guidance !
I'm curious what happened to the generation with this volume normalization issue : did the generation bug with this option on, and therefore no files were available for download, and the audio was generated only after I've ticked it off?
I see only one generation in the history, but it does have yesterday's date though... maybe it's only the generation start timestamp, but it never actually completed ?
meager pike
Sorry that you run into the volume normalisation issues, it is a bit flakey and sometimes stops the download from being generated. We will fix it, but it's not currently worked on. Basically we apply it when you press download - we join each paragraph together and apply the volume normalisation and upload the resulting file, once it's uploaded you can press again and download it. So ticking it off does not result in new generation but new file is "created" by joining existing generations and uploaded.
I've checked the audio, and the issue with the number seems to have been resolved (number normalized skipped) with this approach. Thanks for the guidance !
Great to hear, that pinpoints the issue quite well!
abstract forge
Indeed ! It's so invaluable (necessary actually!) to have an informed interlocutor to make sense of those obscure problems, and to solve them even.
abstract forge
I'll look into adding setting to disable it in paragraph as well later this week...
Hello, thanks again for your attention on that matter ! Has the feature been added ? I've had a look now and didn't find it.
We have an urgent book to produce. If the feature has not yet been added, is there a way to reverse to v1 ?
meager pike
thanks for reminding me! last weeks were a bit crazy with the release of the new editor so it must have slipped through the cracks, let me check and I'll get back to you later today
abstract forge
Thank you for the quick reply ! Looking forward to the update đ
I can't imagine how you guys' weeks must look like đ
meager pike
Change should be live tomorrow afternoon, sorry for the delay!
abstract forge
And it is! Fantastic, thank you!
abstract forge
Hello, I've just tested the workaround again before the delete the test project I had done, but very unfortunately I've noticed it doesn't work perfectly. It's in fact producing even weirder errors, somehow, using the correct Swiss-French form, but reading out the wrong number.
Here's the results I've just had. Every line pairs are : 1. the test line, 2. the result. Generated as a paragraph, with numbers+abbreviations normalization fully off ("skip").
I've put a "->" arrow to signal how the expected number was read out loud, when read wrong :
Et voilà , en français (intro pour donner le contexte de la langue) :
Le nombre de participants est de septante.
70 (septante), ok
On est Ă septante et un kilomĂštres de Lausanne.
71 (septante et un), ok
AprĂšs septante-deux minutes, le fichier Ă©tait prĂȘt.
72 (septante-deux) -> 71 (septante et un)
Le menu coûte septante-trois francs.
73 (septante-trois) -> 71 (septante et un)
Les septante-quatre personnes invitées sont toutes là .
74 (septante-quatre) -> 71 (septante et un)
Si les septante-cinq passagers sont lĂ , nous sommes prĂȘts.
75 (septante-cinq) ->71 (septante et un)
A septante-six, cela commencerait Ă faire beaucoup de monde.
76 (septante-six) ok
DĂšs septante-sept, cela serait trop.
77 (septante-sept), ok
S'il y en a plus de septante-huit, nous pouvons faire deux voyages.
78 (septante-huit), ok
Et voici le dernier chiffre : septante-neuf.
79 (septante-neuf), ok
Good news, the numbers from 80-99 (also different in Swiss vs French) are all read correctly, so only 72-75 are wrong.
80-89
On continue avec huitante.
Puis l'on poursuite avec huitante et un.
Qui est suivi de huitante-deux.
Qui vient avant huitante-trois.
Huitante-quatre se place ensuite.
Huitante-cinq vient aprĂšs.
Arrive ensuite huitante-six.
Nous avons presque fini, avec huitante-sept.
Quelques lignes encore, c'est le tour de huitante-huit.
Suivi de huitante-neuf en avant-dernier.
90-99
Et, enfin!, nonante.
Mais continuons avec nonante et un.
Il a nonante-deux repas.
Les coureurs sont au nombre de nonante-trois.
C'est nonante-quatre personnes qui participaient.
DĂšs nonante-cinq la masse devient critique.
AprĂšs nonante-six, il va falloir faire des adaptations.
S'il y en a nonante-sept, c'est sans doute trop.
Nonante-huit est une limiteâŠ
Nonante-neuf c'est clairement trop !
meager pike
as a sanity check did you regenerate the paragraph after turning the text normalization off?
abstract forge
Yes, with normalization on it's way worse đ , as discussed at the start of this thread
So, I did see the positive changes of "skip normalization", but not for this 72-75 range. Were they absent in the training set?, that would be oddly specific đ
NB : if I write them in rough phonetic, eg "septentedeu" for "septante-deux", the result is good (which didn't with normalization on, which recognized the number anyway and read it wrong).
meager pike
this is oddly specific! it's more like the training set had these number labeled in a wrong way because it's reading them as a completely different numbers
abstract forge
Ok, you confirm my hunch ! Does this mean we can't expect a fix soon and we have to rely on the phonetic approach ?
meager pike
just to not mislead you, I'm not part of our Research team that trained these models, so I don't know if that's really the case, but I cannot really see any other in the app itself and multilang v1 is quite old model, so I think you need to rely on the phonetic approach unfortunately
abstract forge
I wish we could use v2, but last time I checked, about 3 weeks ago, the voices were producing way too many random artifacts, needing way to much post editing to trim them (voice going "whup" or hallucinating a couple of made up words or even whole gibberish sentences at the end of sentences). In the context of french texts.
There's also a glitch in the titles, at least until the last time I've checked about 1 week ago : ebook import sets html titles as titles in Projects, and I've noticed that texts set as title, specially if they are short (but still a full sentence, providing language context, not just one/two words) would result consistently into hallucinations at the end of the sentence (sometime whole giberrish sentence). When I remove the title format on the faulty sentence and regenerate, the hallucination goes away. So we've been removing title formatting on all chapters (lots of clicks đ ) for months to avoid this. Still with v1 in French, for context.
meager pike
We actually released a fix for glitches in titles (or more precisiely breaks after titles), it shouldn't happen anymore, if you notice is again please let me know.