OpenASR/parakeet-tdt-0.6b-v3

Parakeet TDT 0.6B v3 · OpenASR

NVIDIA's multilingual speed flagship — 25 European languages, TDT transducer with native word timestamps

Native speech-to-text in the OpenASR runtime — engineered for peak performance on CPU & GPU, no Python at inference time.

---

✨ Highlights

• 🇪🇺 25 European languages — one 0.6B checkpoint covers bg/hr/cs/da/nl/en/et/fi/fr/de/el/hu/it/lv/lt/mt/pl/pt/ro/sk/sl/es/sv/ru/uk, auto-detecting the spoken language with no prompting
• ⚡ Built for throughput — the Token-and-Duration Transducer (TDT) decoder predicts how many frames each token spans and skips ahead, cutting decode steps versus conventional transducers
• 🕑 Native word timestamps — the duration head is the model's own alignment output, so word-level timings come from the model rather than a uniform approximation
• ✍️ Punctuation and capitalization — trained on transcripts that preserve both (Granary + NeMo ASR Set 3.0; ~670k hours), so the raw decode reads like text
• 🦀 Native in OpenASR — .oasr packs run with no Python at inference, engineered for peak performance on CPU & GPU

🚀 Quickstart

# 1. Install the OpenASR CLI  ·  https://openasr.org
# 2. Pull a build (pick a quant — see the table below)
openasr pull parakeet-tdt-0.6b-v3:q8

# 3. Transcribe
openasr transcribe audio.wav --model parakeet-tdt-0.6b-v3

All builds for this model:

openasr pull parakeet-tdt-0.6b-v3:fp16
openasr pull parakeet-tdt-0.6b-v3:q8
openasr pull parakeet-tdt-0.6b-v3:q4

📦 Available builds

Quant	File (.oasr)	Size	RAM peak	RTF · M1 CPU	RTF · M1 GPU	JFK ΔWER vs fp16
fp16	parakeet-tdt-0.6b-v3-fp16.oasr	1.42 GB	1.59 GB	0.14×	0.05×	0.0%
q8_0	parakeet-tdt-0.6b-v3-q8_0.oasr	919 MB	1.12 GB	0.13×	0.05×	0.0%
q4_k	parakeet-tdt-0.6b-v3-q4_k.oasr	655 MB	920 MB	0.11×	0.05×	0.0%

_{RTF = real-time factor on the fixed 11s JFK clip (lower is faster); RAM peak measured per pack in an isolated subprocess. JFK ΔWER compares each quantized build's JFK transcript to this model's fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy. q8_0 is the recommended default — near-reference quality at a fraction of the footprint.}

🧠 About Parakeet TDT 0.6B v3

Parakeet TDT 0.6B v3 is NVIDIA's 600M-parameter multilingual speech-recognition model, extending the English-only v2 to 25 European languages using the Granary multilingual corpus (~660k hours of pseudo-labeled speech plus ~10k hours of human transcriptions). Architecturally it pairs a FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder: each decode step predicts both the next token and how many 80 ms frames it occupies, then skips ahead by that duration — which is what makes the family a speed benchmark for local transcription, and what gives it model-native word timestamps. The model auto-detects the spoken language, preserves punctuation and capitalization, and is released under the permissive CC-BY-4.0 license. This OpenASR repo repackages the weights as .oasr packs that run natively in the OpenASR runtime — no Python at inference time, all decoding local. The q8_0 build is the recommended default; q4_k is the smallest build for tight-memory devices and fp16 is for maximum fidelity or verification.

⚙️ How these packs were made

Converted from nvidia/parakeet-tdt-0.6b-v3 with the OpenASR importer:

openasr model-pack import parakeet-tdt <src> <out>.oasr \
  --package-id parakeet-tdt-0.6b-v3 --quantization {fp16,q8-0,q4-k}

The .oasr container is GGUF-backed; packs use zero-copy mmap weight binding and graph buffer reuse to keep peak memory low.

⚖️ License

These packs inherit the upstream model's license: CC-BY-4.0 (source). OpenASR packaging retains the upstream copyright and NOTICE; the only modifications are format conversion and quantization.

🙏 Acknowledgements

This pack is a redistribution of parakeet-tdt-0.6b-v3, created and released by NVIDIA (nvidia/parakeet-tdt-0.6b-v3), trained with the NeMo toolkit on the Granary corpus and NeMo ASR Set 3.0. All credit for the architecture (FastConformer encoder, TDT decoder), training, and weights belongs to NVIDIA; the license is inherited from and identical to the upstream model (CC-BY-4.0, which requires attribution — keep this credit when redistributing). OpenASR only performs format conversion, quantization, runtime verification, and local-inference adaptation.

🔗 Links

• 🦀 OpenASR — https://github.com/QuintinShaw/openasr
• 🌐 Website — https://openasr.org
• 🤗 Upstream model — nvidia/parakeet-tdt-0.6b-v3