First time going to a big gym in the UK, and somehow the practice of saying a little “sorry” as you go past someone cracks me up in that setting.

Fill in the blank: "My p-value is smaller than 0.05, so..." Wrong answers only.

I don't download new HF models often, but when I do, it's during the 0.008% of downtime :(

TTSDS2 is one of the papers accepted by the @neuripsconf.bsky.social area chairs but but rejected by the senior area chairs with no explanation as to why. A bit frustrating after the long review process.

Accents are also best seen as a distribution, not a group of labels imo. We tried to incorporate some proxy of accent in TTSDS2, but a simple phone distribution did not work all that well, probably because it’s hard to disentangle from lexical content…

It's been a great #interspeech2025! I presented a TTS-for-ASR paper: www.isca-archive.org/interspeech_... And one on prosody reps: www.isca-archive.org/interspeech_... There were many interesting questions & comments - if you have more and didn't get the chance feel free to send me a message.

I’ll will be presenting this tomorrow at 8.50 at #interspeech2025, come by if you’re interested in prosodic representations!

In other news — if you’re an early bird and at #interspeech, feel free to drop by my poster presentation on scaling synthetic data tomorrow - who doesn’t want to chat about neural scaling laws early in the morning! App: interspeech.app.link?event=687602... Paper: www.isca-archive.org/interspeech_...

A highlight at #interspeech so far: the “hear me out” show&tell in which you can check how the spoken language model Moshi responds based on if it’s your voice or a voice converted version to the opposite gender. Check it out here shreeharsha-bs.github.io/Hear-Me-Out/ 1/2

If you’re interested in ASR for low resource languages, come by at 14.30 in Poster Area 09 at #interspeech today! I’ll be presenting this paper by Ondrej Klejch et al. arxiv.org/abs/2506.04915

Looking forward to present a bunch of things at #INTERSPEECH and #SSW - will put the details here once my thesis final draft is done, which will probably be on the plane to Rotterdam.

One day until the Q2 ttsdsbenchmark.com update. We‘ll see which TTS system tops the leaderboard this time - some new ones have been added that could shake things up.

Followed your advice and can confirm “Ughaaaghaghaa” was my reaction as well.

This figure motivated a lot of my PhD (or at least nudged me into a direction) -- check out arxiv.org/abs/2110.11479 (Hu et al.) if you haven't come across it before, it really frames the problem of synthetic/real speech distributions well.

Spotted a Norwegian flag across the Firth of Forth, didn’t know Norwegians had hytte on this side of the North Sea as well!

More details on this soon! Also this weekend is the last chance to submit your TTS system for the next round of evaluation (Q2 2025) by either messaging me at [email protected] or requesting a model here: huggingface.co/spaces/ttsds...

It’s amazing how a days work can stretch out over a fortnight, and a week of work can be compressed into 24 hours sometimes…

I wonder if there are naturally left-curling and right-curling cats, or if all cats curl both ways.

I’ve only really encountered people trying to avoid sounding like AI… but it makes sense that it would alter how people speak if they interact with it a lot. Makes me sad though since it pushes people towards the mean, which is always the most boring.

Pro tip #1: don’t use a poster tube when travelling to and a from conferences, people might come up to you and ask about your research. Pro tip #2: get a poster tube if you don’t want to talk about your research anymore once you’re done with the conference.

Taylor 2009: „Sometimes the question is raised as to whether we really want a TTS system to sound like a human at all.“ me: I wonder where this is going later: „no matter how good a system is, it will rarely be mistaken for a real person, and we believe this concern can be ignored.“ me: oh no

Didn’t think I’d see myself on Bluesky today! Has been a fun conference so far, anyone interested in what I have in the works make sure to come by my poster tomorrow ;)

Can’t wait for the whale IPA chart.

🧪 SSL (self-supervised learning) models can produce very useful speech representations, but what if we limit their input to prosodic correlates (Pitch, Energy, Voice Activity)? Sarenne Wallbridge and I explored what these representations do (and don’t) encode: arxiv.org/abs/2506.02584 1/2

Why is @overleaf.com down so close to the @neuripsconf.bsky.social deadline 😓 guess I’ll have to take a break now #AcademicSky

When future archeologists dig up the remains of my thesis in 3,000 years.

The average lifespan of a web page is 100 days, and the average lifespan of a website is only 2.7 years. Online content is often at greater risk than older analogue content. #ndf25

In the same year the first German rap song with mainstream appeal was released, American hip hop had already advanced to this level ⬇️

An infinite number of mathematicians walk into a bar. The first orders a beer, the second half a beer, the third quarter of a beer, and so on. The bartender pours two beers and says, "You guys oughta know your limits." (from: youtu.be/1qbiCKrbbYc)

Fixed tag #AcademicSky

I believe it is important to inform the communities that "produce" training data for AI, so I shared on the LibriVox forum how their audio recordings are used for Text-to-Speech research. forum.librivox.org/viewtopic.ph... Here are some takeaways from the discussion that followed 🧵 #AcademySky 🧪 1/7

Sometimes it’s nice to work in a field (Text-to-Speech) where incentives to do well on a leaderboard aren’t as high as for LLMs. Pretty sure ttsdsbenchmark.com is a good measure… for now

Always a joy when a TechCrunch article that cites a paper but makes up a new term for the phenomenon makes it to the Reddit frontpage. And people continue to extrapolate the controlled experiments to what happens in the wild. However training on synthetic data is problematic when… 1/2

something that I'm proud of! I co-wrote this with Ari Sanchez and @ninamarkl.bsky.social because I got mad at the start of my PhD when I started reading up-to-date papers on speech science/technology and they said things like "the two genders" 😂🥲🫠 doi.org/10.1109/SLT6...

We created Approximate Likelihood Matching, a principled (and very effective) method for *cross-tokenizer distillation*! With ALM, you can create ensembles of models from different families, convert existing subword-level models to byte-level and a bunch more🧵

Modern *autoregressive* TTS systems are really hard to debug. They might output audio that has reasonable duration, mel spectrogram, etc. but is just gibberish.

With the market & AI mythos reeling post-DeepSeek, it seems like a good time to reup this year-old paper, on the perils of the "bigger is better" approach to AI, coauthored with @sashamtl.bsky.social + @gaelvaroquaux.bsky.social arxiv.org/abs/2409.14160

Debunking can be a fun way to find an audience in science communication, but I think it can be over-relied on. We should be celebrating more than we are defending, not just because it’s less miserable, but also because it’s good marketing!

Fascinating that the latest text-to-speech models are still going into two (orthogonal?) directions. On one side we have non-autoregressive, efficient models like Kokoro that won’t be great at voice cloning (or not support it at all) or huge billion parameter language model style ones like Llasa.

when the Bluetooth headphones lag behind the video

Planning to submit to @interspeech.bsky.social 2025!? Our author kit is available! Good luck with the 📝!

I’m told it is mandatory in Norway to leave the city and go to a hytte in thewoods on the weekend, so doing my best.

Can't wait for this to be the topic of choice at all my upcoming dinner parties 🫣 arxiv.org/abs/2412.04984

What pitch prediction would you use for in-the-wild speech data (e.g. from YouTube)? CREPE seems to widely used, but algorithmic methods like DIO, Harvest or pYIN might be more robust for data with background noise/etc.

So you know how we often have a mel spectrogram in a system diagram as the input… I usually ask a friend if they want to be in my next paper or presentation when it comes to this.

It looks like deepfake detection/MOS prediction don’t generalise to new TTS systems released in 2023/24. Sure, if you stay within a class of models you have a chance (e.g. NAR + diffusion) but even there I highly doubt that a new model can be detected if hyper params are different. 🧵 on this soon

The starter pack I’ve been looking for!

Hey #linguistics, I've created a starter pack for phoneticians, speech people, and phriends. If you feel like I've left someone out or you have suggestions, just lmk! I want more spectrograms on my feed so let's make that happen. go.bsky.app/LPrwcbq

#AcademicSky #AI #ML #TTS

There are very good speech synthesis models out there now, as everyone should have noticed at this point. These models have been subjectively (so using human judges) evaluated for a long time. However, it's unclear which method will be used going forward - CMOS or SMOS 🧪🧵 (1/9)