Hubert Siuzdak

research @ElevenLabs

[Hubert Siuzdak @HubertSiuzdak]

It's been super fun pushing what's possible with on-device TTS with such a great team. So proud of what we've built - and sharing it at our Summit in Warsaw, where ElevenLabs began, feels pretty special. 🇵🇱

[ElevenLabs @ElevenLabs]

At the ElevenLabs Summit in Warsaw, we previewed on-device Text to Speech - a new model architecture that delivers human-level quality on limited hardware without an internet connection.

[ElevenLabs @ElevenLabs]

Introducing Dubbing v2, our revolutionary new dubbing model.

For the first time, the emotion and performance of the original content is carried over into every language.

[Mati Staniszewski @mati]

So excited to welcome the Papla team on board at @ElevenLabs to help accelerate our work on the best voice interfaces - elevating model quality, building on-device, and scaling forward-deployed engineering.

@dabkowski_piotr and I first met @HubertSiuzdak four years ago (!) at one of the first conferences we attended: Interspeech. He later went on to co-found Papla with three amazing engineers - Dominik, Jakub, and Tomasz - also working toward the future of voice & voice agents.

When we first met, Hubert was presenting his breakthrough on latent speech representation with WavThruVec. Beyond the impact and clarity of the research, one thing was immediately clear: passion. Even through poor photo quality and Covid masks, you can see the happiness and energy beaming through. That same energy extended to his co-founders and the company they built - a team driven by excellence, working side-by-side with customers to bring voice agents to enterprises.

Our Poland team is now more than 50 people strong, as we continue accelerating our presence here.

Dominik, Hubert, Kuba, Tomek - a pleasure to learn from you and work together!

[Hubert Siuzdak @HubertSiuzdak]

@JulienBlanchon thank you! yes

[Hubert Siuzdak @HubertSiuzdak]

I'm joining ElevenLabs, together with the founding team of Papla Media, to continue our shared mission of advancing voice interfaces. Excited for what's ahead.

[Hubert Siuzdak @HubertSiuzdak]

@matiii @ElevenLabs @dabkowski_piotr didn't expect this photo 😅 Mati, Piotr - really appreciate the opportunity. Excited to build together 🇵🇱

[Hubert Siuzdak @HubertSiuzdak]

Cool work and a nice read. But why we rebranded masked language models to diffusion language models 😭

[Jinjie Ni @NiJinjie]

Token crisis: solved. ✅

We pre-trained diffusion language models (DLMs) vs. autoregressive (AR) models from scratch — up to 8B params, 480B tokens, 480 epochs.

Findings:
> DLMs beat AR when tokens are limited, with >3× data potential.
> A 1B DLM trained on just 1B tokens hits 56% HellaSwag & 33% MMLU — no tricks, no cherry-picks.
> No saturation: more repeats = more gains.

🚨 ”https://t.co/jmUcE1kywW”
We also dissected the serious methodological flaws in our parallel work “Diffusion Beats Autoregressive in Data-Constrained Settings” — let’s raise the bar for open review!

🔗 Blog & details:
https://t.co/sEQvYUxElj

18 🧵s ahead:

[Hubert Siuzdak @HubertSiuzdak]

Papla Voicebot is genuinely fun to talk to! Our Papla P1 engine, when running locally, responds in 50-150 ms and it’s also pretty expressive. The main bottleneck right now is Gemini 2.0 Flash latency: the first chunk we pass to TTS typically arrives in ~600 ms. We're exploring a few ideas to make it even snappier, including running a local open-source LLM.

[Papla Media @papla_media]

Here's @HubertSiuzdak having a quick chat with our AI voicebot. It listens, responds in real time, and even laughs - just like a real person.

[Hubert Siuzdak @HubertSiuzdak]

Ok but doing modeling @runwayml is way funnier, feel free to steal

[Hubert Siuzdak @HubertSiuzdak]

Or we just say we're doing modeling now

[Elon Musk @elonmusk]

This false nomenclature of “researcher” and “engineer”, which is a thinly-masked way of describing a two-tier engineering system, is being deleted from @xAI today.

There are only engineers.

Researcher is a relic term from academia.

[Papla Media @papla_media]

Building the future of AI voice…
Allergy season meets beta testing and apparently, our voicebot has opinions now. 🤧🤖
Still more emotionally aware than some customer service lines.
Big things are coming from Papla Media 💛

#ConversationalAI

[Hubert Siuzdak @HubertSiuzdak]

We raised $360k in a pre-seed round - grateful to everyone who believed in us early.

We've got some great things coming, and can't wait to share them soon!

[Papla Media @papla_media]

We're happy to share that Papla Media raised a $360k pre-seed round.
https://t.co/uRa8yI685c

[Papla Media @papla_media]

🚀 New TTS Playground Now Live at Papla Media!
We just launched a powerful new playground on the Papla Media platform designed to make working with AI voices faster and easier than ever.

Now you can:

📝 Type text and hear it instantly in ultra-realistic voices
🎙️ Try multiple voices before deciding which one to use in your app
🔊 Prototype and download audio
🧬 Generate speech from your cloned voices directly via the UI
⚡ Move and experiment fast

Whether you’re building tools, testing ideas, or just playing with voice, this space is for you.

[Papla Media @papla_media]

🚀 Introducing the New Voice Tab
We just made it easier (and more fun) to explore, test, and create voices like never before.

🎧 Browse and preview voices across styles, accents, and tones
🧬 Clone your own voice in a few clicks from just 10 seconds of audio.

Perfect for creators, developers, brands, and anyone building with voice
👉 Jump in, explore the voices, and try cloning your own: https://t.co/0ZFohrqPwl

[Hubert Siuzdak @HubertSiuzdak]

@laion_ai yeah, you can steer emotions/styles with different audio prompts!

[Hubert Siuzdak @HubertSiuzdak]

is this the state of the art in instant voice cloning? What should we compare it to?

[Papla Media @papla_media]

🎙️ Voice Cloning with Papla P1 Just Got Real
With only 10 seconds of audio, P1 can create a highly realistic voice clone that captures tone, accent, and personality.

This means fast, scalable voice personalization that sounds authentic and deeply human.

🔹 Just 10 seconds of voice input

🔹 Natural rhythm and emotional nuance

🔹 Ready for content, games, apps, and more

From hyper-personalized experiences to next-gen audio production, Papla P1 voice cloning opens up endless possibilities.

🔊  Let us know what you think in the comments, we’d love to hear your thoughts.

[Papla Media @papla_media]

We're live on the TTS Arena! Check out Papla P1 and see how it compares → https://t.co/6WF60dOVj3

[Hubert Siuzdak @HubertSiuzdak]

Today we're launching our platform with Papla P1 - our ultra-realistic text-to-speech model, available via API!

Create truly conversational voices - and with real-time streaming, it's perfect for live, interactive apps.

[Papla Media @papla_media]

Introducing Papla P1 and our real-time API for developers! 🚀

Papla P1 is our advanced text-to-speech model, now available through a developer-friendly platform. Easily generate realistic speech, clone voices, and build natural-sounding conversations into your apps.

[Hubert Siuzdak @HubertSiuzdak]

Excited to be at NeurIPS 2024 in beautiful Vancouver! Come check out our poster on SNAC at the Saturday Audio Workshop. Big thanks to Luca @lucalanze and Florian for the collab. DM me if you’d like to chat about audio generative models, conversational AI, or startups!

[arXiv Sound @ArxivSound]

``SNAC: Multi-Scale Neural Audio Codec,'' Hubert Siuzdak, Florian Gr\"otschla, Luca A. Lanzend\"orfer, https://t.co/GQrDlEaHeQ

[Hubert Siuzdak @HubertSiuzdak]

@jpclap see you there!