Daan van Esch
@daanvanesch.nl
I work on speech and language technologies at Google. I like languages, history, maps, traveling, cycling, and buying way too many books.
Your books are also having a great time chatting with each other in the linguistics section here at the Kinokuniya bookstore on Orchard Road in Singapore!
November 22, 2025 at 7:12 AM
Your books are also having a great time chatting with each other in the linguistics section here at the Kinokuniya bookstore on Orchard Road in Singapore!
Ik doe tegenwoordig precies hetzelfde: ik dicteer in mijn coding agent. En dat gaat lastiger op kantoor dan thuis op mijn werkkamer. Voor mij is er zowat genoeg verschil in mijn productiviteit ten opzichte van typen om een dagje meer thuis te werken...
October 21, 2025 at 5:34 PM
Ik doe tegenwoordig precies hetzelfde: ik dicteer in mijn coding agent. En dat gaat lastiger op kantoor dan thuis op mijn werkkamer. Voor mij is er zowat genoeg verschil in mijn productiviteit ten opzichte van typen om een dagje meer thuis te werken...
Reposted by Daan van Esch
VarDial 2026 will be colocated with @eaclmeeting.bsky.social! We're looking forward to your papers on NLP for similar languages, varieties and dialects :)
Deadline: Dec 19 (Jan 2 for pre-reviewed ARR papers)
sites.google.com/view/vardial...
Deadline: Dec 19 (Jan 2 for pre-reviewed ARR papers)
sites.google.com/view/vardial...
October 21, 2025 at 10:36 AM
VarDial 2026 will be colocated with @eaclmeeting.bsky.social! We're looking forward to your papers on NLP for similar languages, varieties and dialects :)
Deadline: Dec 19 (Jan 2 for pre-reviewed ARR papers)
sites.google.com/view/vardial...
Deadline: Dec 19 (Jan 2 for pre-reviewed ARR papers)
sites.google.com/view/vardial...
I've been watching Feuer und Flamme season 10 in Heidelberg the last few weeks, which captures the everyday work of the local fire service. Definitely needs dialectal ASR for some speakers!
October 20, 2025 at 6:37 AM
I've been watching Feuer und Flamme season 10 in Heidelberg the last few weeks, which captures the everyday work of the local fire service. Definitely needs dialectal ASR for some speakers!
Reposted by Daan van Esch
The ConlangCrafter pipeline harnesses an LLM to generate a description of a constructed language and self refines it in the process. We decompose language creation into phonology, grammar, and lexicon, and then translate sentences while constructing new needed grammar points.
October 11, 2025 at 5:35 AM
The ConlangCrafter pipeline harnesses an LLM to generate a description of a constructed language and self refines it in the process. We decompose language creation into phonology, grammar, and lexicon, and then translate sentences while constructing new needed grammar points.
That's very cool, do you have a sense of how many different Soviet minority languages there are in this collection?
September 17, 2025 at 6:52 AM
That's very cool, do you have a sense of how many different Soviet minority languages there are in this collection?
Great to see this highly multilingual model: 1,000+ languages!
EPFL, ETH Zurich & CSCS just released Apertus, Switzerland’s first fully open-source large language model.
Trained on 15T tokens in 1,000+ languages, it’s built for transparency, responsibility & the public good.
Read more: actu.epfl.ch/news/apertus...
Trained on 15T tokens in 1,000+ languages, it’s built for transparency, responsibility & the public good.
Read more: actu.epfl.ch/news/apertus...
September 3, 2025 at 6:17 AM
Great to see this highly multilingual model: 1,000+ languages!
Oh wow that is very cool! 😮
September 1, 2025 at 8:56 AM
Oh wow that is very cool! 😮
Sounds great, looking forward!!
August 31, 2025 at 6:29 AM
Sounds great, looking forward!!
Yes I would've loved to come!! I'll have to do another trip to Leeuwarden sometime soon... In the meantime, let me know if you have time to come over for lunch one of these days! I'm in London and Stockholm for a bit this month but October is looking pretty open!
August 30, 2025 at 11:50 AM
Yes I would've loved to come!! I'll have to do another trip to Leeuwarden sometime soon... In the meantime, let me know if you have time to come over for lunch one of these days! I'm in London and Stockholm for a bit this month but October is looking pretty open!
(of course you may want to apply something like voice conversion if it's preferred for the website to be read out by a different speaker)
August 30, 2025 at 7:45 AM
(of course you may want to apply something like voice conversion if it's preferred for the website to be read out by a different speaker)
So cool! I'm sad I wasn't able to make it but it sounds like a great event. The last paragraph is interesting, is their website already translated into Bildts (manually or automatically)? I imagine someone could use something like Cloud Run to make a simple TTS service using one of these models!
August 30, 2025 at 7:45 AM
So cool! I'm sad I wasn't able to make it but it sounds like a great event. The last paragraph is interesting, is their website already translated into Bildts (manually or automatically)? I imagine someone could use something like Cloud Run to make a simple TTS service using one of these models!
Reposted by Daan van Esch
AI efficiency is important. The median Gemini Apps text prompt in May 2025 used 0.24 Wh of energy (<9 seconds of TV watching) & 0.26 mL (~5 drops) of water. Over 12 months, we reduced the energy footprint of a median text prompt 33x, while improving quality:
cloud.google.com/blog/product...
cloud.google.com/blog/product...
August 21, 2025 at 1:39 PM
AI efficiency is important. The median Gemini Apps text prompt in May 2025 used 0.24 Wh of energy (<9 seconds of TV watching) & 0.26 mL (~5 drops) of water. Over 12 months, we reduced the energy footprint of a median text prompt 33x, while improving quality:
cloud.google.com/blog/product...
cloud.google.com/blog/product...
Great! Sent you a private message
August 21, 2025 at 5:29 AM
Great! Sent you a private message
Wait how long are you in town for?
August 20, 2025 at 8:39 PM
Wait how long are you in town for?
Reposted by Daan van Esch
⏳ Just 1 week to go! 🎉
#Interspeech2025 kicks off next week in Rotterdam, the Netherlands 🗣️🌍
We can’t wait to welcome everyone for a week full of talks, posters, workshops & networking.
📅 See you soon!
Comment below, are you joining? 🥰
#Interspeech2025 kicks off next week in Rotterdam, the Netherlands 🗣️🌍
We can’t wait to welcome everyone for a week full of talks, posters, workshops & networking.
📅 See you soon!
Comment below, are you joining? 🥰
August 10, 2025 at 11:14 AM
⏳ Just 1 week to go! 🎉
#Interspeech2025 kicks off next week in Rotterdam, the Netherlands 🗣️🌍
We can’t wait to welcome everyone for a week full of talks, posters, workshops & networking.
📅 See you soon!
Comment below, are you joining? 🥰
#Interspeech2025 kicks off next week in Rotterdam, the Netherlands 🗣️🌍
We can’t wait to welcome everyone for a week full of talks, posters, workshops & networking.
📅 See you soon!
Comment below, are you joining? 🥰
Reposted by Daan van Esch
🥳 Happy to open up the registrations for the CLIN conference! You can find more information here: clin35.ccl.kuleuven.be/registration The website has also been updated with more information for the presenters, with a programme, and with information about the venue. See you soon at #CLIN35!
August 8, 2025 at 1:14 PM
🥳 Happy to open up the registrations for the CLIN conference! You can find more information here: clin35.ccl.kuleuven.be/registration The website has also been updated with more information for the presenters, with a programme, and with information about the venue. See you soon at #CLIN35!
Today the NRC Handelsblad (one of the Dutch newspapers of record) had a lengthy article on this: www.nrc.nl/nieuws/2025/... Somehow I remembered this post (it's an important question I think for our democracies) so I figured I'd send you the link, let me know if you have paywall trouble!
Chatbots hallucineren nogal eens, dat is een risico als je ze om stemadvies vraagt
Stemhulpen: Veel kiezers zullen voor de Tweede Kamerverkiezingen bij chatbots als ChatGPT aankloppen voor stemadvies. Is dat wel verstandig?
www.nrc.nl
August 3, 2025 at 8:35 PM
Today the NRC Handelsblad (one of the Dutch newspapers of record) had a lengthy article on this: www.nrc.nl/nieuws/2025/... Somehow I remembered this post (it's an important question I think for our democracies) so I figured I'd send you the link, let me know if you have paywall trouble!
This is interesting, I hadn't seen this, thanks! It does suggest, I suppose, that as consumer hardware becomes more powerful, that a 100B and eventually a 1T model eventually also gets called a small language model, I think? That may be right at that time, but it's a little unsatisfying to me.
July 22, 2025 at 6:11 AM
This is interesting, I hadn't seen this, thanks! It does suggest, I suppose, that as consumer hardware becomes more powerful, that a 100B and eventually a 1T model eventually also gets called a small language model, I think? That may be right at that time, but it's a little unsatisfying to me.
I always wonder about this kind of terminology: what's large today will be small in a few years, just like this smartphone that I'm typing on is far more powerful than mostly any computer of the 1990s, yet we obviously don't call it a supercomputer. Two years ago 10B would've no doubt been an LLM.
July 22, 2025 at 6:09 AM
I always wonder about this kind of terminology: what's large today will be small in a few years, just like this smartphone that I'm typing on is far more powerful than mostly any computer of the 1990s, yet we obviously don't call it a supercomputer. Two years ago 10B would've no doubt been an LLM.
e.g. [2502.02737] SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model share.google/cmaS3DYNTB2X... but to be honest it's always struck me as a bit unusual, like isn't Gemma 4B an LLM? Is a 200B n-gram statistical model an LLM just based on its size?
SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model
While large language models have facilitated breakthroughs in many applications of artificial intelligence, their inherent largeness makes them computationally expensive and challenging to deploy in r...
share.google
July 21, 2025 at 8:34 PM
e.g. [2502.02737] SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model share.google/cmaS3DYNTB2X... but to be honest it's always struck me as a bit unusual, like isn't Gemma 4B an LLM? Is a 200B n-gram statistical model an LLM just based on its size?
Reposted by Daan van Esch
Hi #NLP community, I'm urgently looking for an emergency reviewer for the ARR Linguistic Theories track. The paper investigates and measures orthography across many languages. Please shoot me a quick email if you can review!
June 21, 2025 at 10:34 AM
Hi #NLP community, I'm urgently looking for an emergency reviewer for the ARR Linguistic Theories track. The paper investigates and measures orthography across many languages. Please shoot me a quick email if you can review!