Comment by spudlyo
12 hours ago
https://fabulae.orbilii.com - A Latin language audiobook containing both volumes of A.D. Godley's "The Fables of Orbilius" which are entertaining intermediate Latin stories written at the turn of the 20th century for use in British schools.
The audio itself is in delivered with the restored classical Latin pronunciation and contains spoken nuances like vowel nasalization, elision, and prodelision. I used a wav2vec2 model to generate forced-alignment data, so you can easily follow along with the text as it's being read aloud. Each word in the text is richly enhanced with pedagogical details like: lemma, citation forms / principal parts, morphological segmentation and analysis, inflectional class with complete declension/conjugation tables, verb category, contextual lexical information (i.e. specific L&S sense), Classical Latin IPA pronunciation, UD style syntactic function and relations, predicate valency, and concise English definitions. Spoken instances of elision and prodelision can also optionally be displayed.
From a nerd perspective, one of the things that makes the system interesting is that the documents themselves are entirely in XML, and I use the soon-to-be-removed-everywhere in-browser XSLT feature to render the HTML, so you can see the underlying XML structure if you view source. Each sentence in the text contains provenance records so you can see the language model and parameters that performed each language enrichment task. The site UI is largely bilingual, and you can view it in either Latin or English.
Woah this is so interesting. Confirmed it myself, you actually implemented elision!!! I really like your idea of capturing provenance in-document as well. This is humans speech, right? If you can do forced alignment on Latin, it gets me wondering if one could use a wav2vec especially turned to convert IPA into speech somewhat deterministically to compensate for how there is not heaps of training data in Latin.
Haha, thank you for using multiple exclamation marks!!! I'm thrilled that someone is as excited as I am about this feature, it made all the time I spent manually correcting/marking all of Gemini's elision hallucinations worthwhile :)
The recordings were made by my friend and Latin tutor. I used an Italian wav2vec2 model fine-tuned on Latin[0], which works pretty well for Latin only content, but falls down with mixed English / Latin. I found that some of the slower, larger Meta models worked fine for this use case though.
Some folks on the Latin language Discord said they got pretty good results using Kokoro TTS and X-Voice for creating synthetic Latin, as you suggested using IPA.
[0]: https://huggingface.co/lsb/wav2vec2-base-it-latin