Skip to content
← Models
Open weightsNVIDIA/10 trustVerified

Parakeet TDT 0.6B v3

NVIDIA's 600M-parameter multilingual transcriber: 25 European languages with auto language detection, word timestamps, and ~20x real-time speed — CC-BY-4.0 open.

Visit official page →

About

Parakeet TDT 0.6B v3 (August 2025, Hugging Face nvidia/parakeet-tdt-0.6b-v3, technical report arXiv 2509.14128) is the multilingual successor in NVIDIA's Parakeet line: a 600M FastConformer encoder with a Token-and-Duration Transducer decoder, expanded from English-only v2 to 25 European languages with automatic language detection — no prompting needed. It emits punctuated, capitalized transcripts with word- and segment-level timestamps and handles up to 24 minutes of audio in a single pass. NVIDIA pairs it with Canary-1B-v2 (the 1B sibling covering ASR plus speech-to-text translation), and the report claims Canary-1B-v2 beats Whisper-large-v3 on English ASR at 10x the speed with competitive multilingual ASR/AST against Seamless-M4T-v2-large. Community CPU benchmarks back the speed story (~2.16% WER on LibriSpeech test-clean at ~20x real-time). The whole thing is CC-BY-4.0, so it's genuinely usable commercially — the go-to open replacement for Whisper in European-language transcription pipelines.

Specs

Context window

Up to 24 minutes of audio in a single pass (full attention); chunked mode extends further

Pricing

Free / open weights (CC-BY-4.0)

Free quota

Fully open — self-host from Hugging Face; GGUF builds for CPU inference

API access

No hosted API — self-host via NeMo toolkit, Hugging Face, or GGUF/ONNX community builds

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Canary-1B-v2 outperforms Whisper-large-v3 on English ASR at 10x speed (author-reported, arXiv 2509.14128)\n- Community CPU benchmark: 2.16% WER on LibriSpeech test-clean, ~0.05 RTF (independent, small-scale)\n- Competitive multilingual ASR/AST vs Seamless-M4T-v2-large (author-reported)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

transcriptionspeech-translationsubtitlingvoice-analyticsaccessibility

Trust score — 8/10

/10 trustVerified

Scored 8/10 (TIER verified, metrics-v1). Evidence: NVIDIA technical report (arXiv 2509.14128); open weights on Hugging Face; independent community CPU benchmarks confirming the speed claims. Strengths: CC-BY-4.0 commercial license, 25 languages with auto-detection, timestamps, genuinely fast. Docked for: (1) translation (AST) lives in the Canary sibling, not this checkpoint; (2) European-languages-only coverage; (3) headline Whisper-beating figures are NVIDIA's own. License: CC-BY-4.0.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.0/10
Maintenance · recency, releases, issue load
9.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · 600M FastConformer-TDT; 25 European languages, auto language detection, word timestamps
  • · Released 2025-08-14 on Hugging Face; part of the Granary series with Canary-1B-v2
  • · Headline Whisper-beating figures are NVIDIA's own; community CPU benchmarks confirm the speed story (~2.16% WER LibriSpeech test-clean, ~20x real-time)
  • · Speech-to-text translation lives in the Canary-1B-v2 sibling, not this checkpoint

Sources