Another rainy day

Posted by Victor Mair

https://languagelog.ldc.upenn.edu/nll/?p=72894&utm_source=rss&utm_medium=rss&utm_campaign=unifying-arabic-topolects-through-ai

https://languagelog.ldc.upenn.edu/nll/?p=72894

Meet Habibi – the Chinese AI uniting 20 Arabic dialects in a Middle East first
Lead author says there are many differences between Arabic dialects and Modern Standard Arabic, which is used in official circumstances
Zhao Ziwen, SCMP, 28 Feb 2026

The paper that presents this new model is called “Habibi: Laying the Open-Source Foundation of Unified-Dialectal Arabic Speech Synthesis”. It was published last month on arXiv, an open-access repository that is not peer-reviewed.  I will be interested to hear what Language Log readers think of its prospects.

Led by Shanghai Jiao Tong University’s X-LANCE Lab – one of China’s top audiovisual and language processing research entities – the model is named Habibi, meaning “my dear” in Arabic.

In presenting their findings, the research team spearheaded by Chen Yushen described the project in a paper as “the first open-source framework for unified-dialectal Arabic speech synthesis”.

They introduce a concept that is new to me:   "zero-shot".

Habibi has the “zero-shot” ability, meaning the model can easily clone a voice by using just a short reference audio clip, without prior explicit or extensive training. This allows applications in highly efficient and on-the-fly scenarios.

According to Wikipedia,

Zero-shot learning (ZSL) is a problem setup in deep learning where, at test time, a learner observes samples from classes which were not observed during training, and needs to predict the class that they belong to. The name is a play on words based on the earlier concept of one-shot learning, in which classification can be learned from only one, or a few, examples.

Zero-shot methods generally work by associating observed and non-observed classes through some form of auxiliary information, which encodes observable distinguishing properties of objects.  For example, given a set of images of animals to be classified, along with auxiliary textual descriptions of what animals look like, an artificial intelligence model which has been trained to recognize horses, but has never been given a zebra, can still recognize a zebra when it also knows that zebras look like striped horses. This problem is widely studied in computer vision, natural language processing, and machine perception.

Selected readings

• "LLMs and tree-structuring" (9/18/25)
• "Radial dendrograms (7/26/23)
• "Language trees and script trees" (12/27/21)
• "AMI not AGI?" (8/2/25)

Addendum

In case you're interested, "Habibi" itself is an Arabic word worth learning in one of its 20 plus topolects:  Syrian, Egyptian, Jordanian, Levantine….   Because of its wide range of meanings, nuances, and usages, be careful of how, when, and to whom you use it.

Listen here.

[Thanks to Mark Metcalf]

https://languagelog.ldc.upenn.edu/nll/?p=72894&utm_source=rss&utm_medium=rss&utm_campaign=unifying-arabic-topolects-through-ai

https://languagelog.ldc.upenn.edu/nll/?p=72894

Posted by Mark Liberman

https://languagelog.ldc.upenn.edu/nll/?p=72892&utm_source=rss&utm_medium=rss&utm_campaign=whats-still-wrong-with-ai-text-to-speech

https://languagelog.ldc.upenn.edu/nll/?p=72892

Text-To-Speech technology has improved enormously over the decades — but there's still some headroom, as Lane Greene has recently underlined for me, expressing dissatisfaction with the AI-read versions of digital articles at The Economist magazine:

I downloaded a handful of "AI Narrated" stories (as the Economist calls then), and then the human-read versions for the ones that made it into print. Before getting to Lane's complaint about repetitive prosody, I noticed a few (minor) old-fashioned errors, such as this parsing (or interpretation?) problem that makes it sound like a Supreme Court tariff is ruling rains (?) inside of Donald Trump:

Or this focus problem, where the human reader helpfully contrasts dollars with euros,

…which the AI narrative failed to do:

As for the stereotyped pitch accents that Lane complained about, here's one of the first sentences in the AI version of the example story that Lane sent me (that link will send you to the slightly-revised print version):

As he observed, it sounds fine. The print version has modified the text somewhat, but you should be able to hear that the corresponding phrase deploys a more varied set of pitch accents:

We can zero in on the subject noun phrase to see as well as hear the difference, first in the AI version:

And now the human version:

You can listen to as much as you like of the two versions, and see whether you agree with Lane that "this high-then-falling curve that is fine in one sentence, but repeated 50 times in a row is awful":

It's easy to quantify Lane's falling-falling-falling perception by looking at syllable-scale dipole statistics, showing a two-dimensional density plot comparing time differences against pitch differences:

AI Reader	Human Reader

There's a lot more to say, and many more articles to look at, but that's enough for today.

https://languagelog.ldc.upenn.edu/nll/?p=72892&utm_source=rss&utm_medium=rss&utm_campaign=whats-still-wrong-with-ai-text-to-speech

https://languagelog.ldc.upenn.edu/nll/?p=72892