Parakeet TDT 0.6B v3

NVIDIA's compact multilingual speech-to-text model: 25 European languages in 0.6B parameters, with timestamps and a CC BY 4.0 license.

In short. Parakeet TDT 0.6B v3 turns recorded speech into punctuated text in 25 European languages. It works out which language is spoken by itself, gives a timestamp for every word, and is small enough for a single modest graphics card or a local C++ app. Commercial use is free as long as you credit NVIDIA.

Summary

Parakeet TDT 0.6B v3 is an NVIDIA speech-to-text model with about 627 million parameters (the card says 600 million). It extends the English-only v2 model to 25 European languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. It detects the language by itself, adds punctuation and capitalization, and returns word-level and segment-level timestamps. The architecture is a FastConformer encoder with a token-and-duration transducer (TDT) decoder, which predicts how long each token lasts as well as the token itself, so decoding can skip over audio frames. The card gives the Hugging Face release date as 14 August 2025, and a technical report appeared in September 2025.

NVIDIA trained it on about 660,000 hours of pseudo-labeled speech from its Granary corpus plus about 10,000 hours of human-transcribed data, then fine-tuned it on roughly 7,500 hours of high-quality human transcripts. Audio must be 16 kHz mono, as .wav or .flac. With full attention, one pass handles recordings of up to 24 minutes on an 80 GB A100, and local attention extends that to about 3 hours. The model runs in NVIDIA's NeMo toolkit and in Hugging Face Transformers, the card shows a chunked streaming script for NeMo, and a lightweight C++ runtime called NeMo-Speech.cpp can use an 8-bit GGUF file of 0.71 GB from the same repository.

NVIDIA reports its own numbers. The average word error rate across the English Open ASR Leaderboard datasets is 6.34%. On the multilingual sets the averages are 11.97% on FLEURS, 7.83% on MLS and 11.98% on CoVoST, using greedy decoding, no external language model, and text with punctuation and capitalization removed. Italian (3.00% on FLEURS), Spanish (3.45%) and Portuguese (4.76%) are among the best results, while Latvian (22.84%), Slovenian (24.03%) and Greek (20.70%) are among the weakest. The card notes that its training data uses European Portuguese while most benchmarks use Brazilian Portuguese. In a noise test with music and background sound at 0 dB signal-to-noise ratio, the English average rises from 6.34% to 11.66%.

The license is CC BY 4.0, which allows commercial use if you credit NVIDIA, link to the license and note any changes. NVIDIA lists Linux and its Ampere, Blackwell, Hopper and Volta GPUs as the supported platform. The card says the model needs at least 2 GB of RAM to load, and that more memory allows longer audio. It is a recorded-speech model first. For live voice agents NVIDIA offers separate cache-aware streaming models, such as Nemotron 3.5 ASR.

Good for

  • Fast, accurate transcription of recordings in 25 European languages, with the language detected automatically
  • Subtitles and searchable transcripts, since every word comes with a timestamp
  • Long recordings, up to 24 minutes in one pass on an 80 GB GPU, or up to about 3 hours with local attention
  • Commercial products, as long as you credit NVIDIA under CC BY 4.0

Watch out for

  • Only the 25 listed European languages are covered, so there is no Chinese, Japanese, Korean or Arabic
  • Accuracy varies a lot by language, from about 3% word error rate in Italian to more than 20% in Latvian, Slovenian and Greek on FLEURS
  • Noise hurts, and NVIDIA's own test shows the English error rate nearly doubling at 0 dB signal-to-noise ratio
  • NVIDIA supports Linux on its own GPUs, and the card warns that words outside the training vocabulary are unlikely to be recognized
  • All word error rates are NVIDIA's own

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Parakeet TDT 0.6B v3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Parakeet TDT 0.6B v3 FP32 627M 2.51 GB 2.89 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

A Creative Commons license that lets anyone copy, share and adapt the work for any purpose, including commercial use, as long as they give credit. It was written for creative and data works rather than software, but some speech and data releases use it for model weights.

  • Attribution or model name must be shown

More about CC BY 4.0

See also

Frequently asked questions

What is Parakeet TDT 0.6B v3?
Parakeet TDT 0.6B v3 turns recorded speech into punctuated text in 25 European languages. It works out which language is spoken by itself, gives a timestamp for every word, and is small enough for a single modest graphics card or a local C++ app. Commercial use is free as long as you credit NVIDIA.
Can I use Parakeet TDT 0.6B v3 commercially?
Yes. Parakeet TDT 0.6B v3 is released under CC BY 4.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Parakeet TDT 0.6B v3 open source?
Parakeet TDT 0.6B v3 is open-weight, not open source in the strict sense. You can download the model files, but the CC BY 4.0 adds conditions that the Open Source Initiative definition does not allow.
Can I run Parakeet TDT 0.6B v3 on my own computer?
Yes. The smallest variant, Parakeet TDT 0.6B v3, needs about 2.89 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Parakeet TDT 0.6B v3 released?
Parakeet TDT 0.6B v3 was released by NVIDIA on August 14, 2025.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Nemotron 3.5 ASR NVIDIA 638M – OpenMDW-1.1 (Permissive) 2.93 GB Laptop or phone
Granite Speech 5.0 IBM 470M – 2.31B – Apache 2.0 (Permissive) 1.09 GB Laptop or phone
Cohere Transcribe Cohere 2.07B – Apache 2.0 (Permissive) 4.75 GB Laptop or phone
Whisper large-v3-turbo OpenAI 810M – 1.54B – MIT (Permissive) 6 GB Laptop or phone