Piotr Żelasko

AI + Speech @ Nvidia. PhD @ AGH-UST, ex-JHU. My interests: speech processing technology; ML/AI software engineering. Building OSS for Speech AI.

[Piotr Żelasko @PiotrZelasko]

@RonanKMcGovern I don’t believe that we did at the time.
Comparing WER across languages is not straightforward, some have structurally higher levels

[Piotr Żelasko @PiotrZelasko]

Second big release from us today: Nemotron-3.5-ASR-Streaming!
🌎40 languages
⚡️80ms - 1s controllable latency
🔥240 - 2400 concurrent streams on 1xH100
🧱FastConformer Cache-Aware RNN-T architecture

https://t.co/lxmcAnKeOl

[Jensen Huang @JensenHuang]

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.

AI will transform every industry, power every company, and be built by every country.

Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.

The world needs both frontier closed models and frontier open models.

https://t.co/AUKzoQ5Ikb

[Ettore Di Giacinto @mudler_it]

NVIDIA just dropped Nemotron-3.5-ASR: one 0.6B model, 40+ languages, streaming.

parakeet.cpp already runs it. On a plain CPU, 2.5x faster than @NVIDIAAI 's Nemo runtime, output byte-for-byte identical (WER 0).

No GPU needed. Offline or real-time. Pick a language with --lang, or auto.

GPU numbers are coming to compare with Nemo framework.

[Piotr Żelasko @PiotrZelasko]

@CharlesBanks99 Should be able to handle that if not too extreme

[Piotr Żelasko @PiotrZelasko]

@gabimoncha It should when on auto mode, let us know your experience :)

[Piotr Żelasko @PiotrZelasko]

@MehdiAllahyari @mervenoyann Yes it should work, try our new finetuning skill to get started, it has all the details and will get you to a good starting setup

https://t.co/iRE0DX1cqV

[Piotr Żelasko @PiotrZelasko]

@abdallah_abra You can use LM decoding to achieve that

[Piotr Żelasko @PiotrZelasko]

@TrelisResearch For most languages offline Parakeet is going to be more accurate, offline-streaming WER gap is real.

[Piotr Żelasko @PiotrZelasko]

@acul_SR Generally yeah, usually using a pretrained model is better than random initialization for new language expansion, see our newly added fine-tuning skill for best practices https://t.co/iRE0DX1Kgt

[Pavlo Molchanov @PavloMolchanov]

Nemotron 3 Ultra (550B-A55B) is here - our strongest open-weight model and full training recipe to date.

Heavy emphasis on real-world inference efficiency for long-context agentic workloads.

Everything is open 🤗: base, post-trained, reward checkpoints, NVFP4 quantized versions, training data, and recipes.

Key technical highlights ‼️:
- 550B total / 55B active parameters
- Hybrid Mamba2-Transformer (~4:1 Mamba:Attention)
- Pretrained in NVFP4 on 20T tokens
- LatentMoE architecture
- Two-stage MOPD post-training
- Native MTP

Technical details in the thread 👇

[NVIDIA AI @NVIDIAAI]

Today we're shipping Nemotron 3 Ultra.

A 550B MoE frontier-intelligence open model built for long-running agents.

It delivers 5x faster inference and lowers the cost of complex agentic tasks by up to 30% versus other open frontier models.

[Piotr Żelasko @PiotrZelasko]

Nemotron 3 Ultra is now available to download: model, pre-training data, post-training data, tech report - everything out in the open. Enjoy!

https://t.co/yglYmy1cb5

[Piotr Żelasko @PiotrZelasko]

Nemotron 3 Ultra just announced, built for the agentic era - compared to other open in models in 500B-1T size class:
🔥 5x faster inference
💰 30% lower cost
🤖 Pretty smart with AA index 48
💎 Trained in NVFP4

Release coming up soon!

[Piotr Żelasko @PiotrZelasko]

🍳HF team cooked here!
🔥Voice agent for Reachy Mini powered by our lightning-fast Parakeet TDT model.

[Andi Marafioti @andimarafioti]

Introducing local Reachy Mini conversations: free chats forever!
So fast that we had to hardcode delays to stop it from interrupting you mid-sentence.

We built an open-source Realtime API powered by llama.cpp:
Parakeet -> Gemma 4 E4B -> Qwen3TTS

Run it anywhere you run local LMs. Video shows DGX Spark and a 36GB M3 Pro MacBook.

Blog: https://t.co/3acRikm2HZ

[Piotr Żelasko @PiotrZelasko]

@md_butcher in my tests it's the same +/- noise (0.05); only the LLM decoder is quantized to nvfp4, the audio encoder stays in bf16

[Piotr Żelasko @PiotrZelasko]

Today we released Nemotron-3-Nano-Omni-30B-A3B - our first Omni model, with speech and audio understanding capabilities powered by parakeet-tdt-0.6b-v2 encoder.
🫡1st position on VoiceBench
🌏English only
🎙️5.95% WER on Open ASR Leaderboard
📽️Video+audio understanding

[Piotr Żelasko @PiotrZelasko]

@matthen2 Both larger and smaller directions are interesting, hard to say which one more

[Piotr Żelasko @PiotrZelasko]

@BENSDJAMEL We will expand to more languages. We could have used parakeet v3 here with 25 language support, but since the model was trained with EN only data, it didn't make sense as parakeet v2 is slightly better in EN ASR than v3.

[Piotr Żelasko @PiotrZelasko]

@teortaxesTex Gotta start somewhere!
Tech report coming right up
https://t.co/p5m94k1fnD

[Piotr Żelasko @PiotrZelasko]

Tech report https://t.co/difVTI2OcH

[Piotr Żelasko @PiotrZelasko]

Tech report https://t.co/difVTI2OcH

[Piotr Żelasko @PiotrZelasko]

https://t.co/GAr4eNBSut