Parakeet (NVIDIA)

Parakeet is NVIDIA's family of open-weight speech recognition models, released through its NeMo toolkit. The first Parakeet models came out in January 2024, built in collaboration with Suno.ai. They used CTC and RNN-Transducer decoders at 0.6 and 1.1 billion parameters, were trained on 64,000 hours of audio, topped Hugging Face's Open ASR Leaderboard that month and were licensed CC-BY-4.0. The best-known current models are Parakeet TDT 0.6B v2 (English, May 1, 2025) and v3 (25 European languages, August 14, 2025). They have become the main open alternative to OpenAI's Whisper, especially where transcripts drive captions or editing.

Architecture

Parakeet TDT pairs a FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder. The decoder predicts each token together with how long it lasts, aligned to the audio frames. Word, segment and character timestamps therefore come straight out of the model, and the decoder has less room to invent text where no speech exists than an encoder-decoder model such as Whisper. The 600-million-parameter models handle up to 24 minutes of audio in one pass with full attention on an 80 GB A100. With local attention, v3 handles up to three hours. NVIDIA trained v3 on about 10,000 hours of human-transcribed audio plus 660,000 hours of pseudo-labeled audio from its Granary dataset.

Accuracy and Speed

On the Open ASR Leaderboard's English view (data version October 17, 2026), Parakeet TDT 0.6B v2 averages 6.02% word error rate and v3 6.18%, against 7.16% for Whisper large-v3 and 7.64% for large-v3-turbo. NVIDIA's larger Canary-Qwen-2.5B scores slightly better still, at 5.94%. The speed gap is wider: the leaderboard's inverse real-time factor is about 6,000 for Parakeet, against 470 for Whisper large-v3 and 794 for turbo. Three caveats apply:

• Long-form audio: the models are nearly tied on the leaderboard's long-form tab (v3 13.37%, turbo 13.60%).
• Other languages: Whisper large-v3 beats v3 in German, Italian, Spanish, Portuguese and Dutch. Parakeet's English lead doesn't carry over automatically.
• Noise: v3's model card reports relative accuracy drops of 12% at 10 dB signal-to-noise, 30% at 5 dB and 84% at 0 dB.

NVIDIA has claimed since 2024 that Parakeet is resilient to music and silence. That's a vendor claim, and no independent study of Parakeet hallucination was found. One dated local test did find v3 dropping whole sentences, depending on how audio was chunked, which neither the model card nor the leaderboard shows (see Whisper vs Parakeet). Test v3 on your own long-form audio before relying on it.

Licenses and Runtimes

The TDT 0.6B models are CC-BY-4.0, which allows commercial use but requires attribution to NVIDIA. Newer models don't all follow suit: parakeet-unified-en-0.6b (April 2026), a streaming model, uses the NVIDIA Open Model License, so check each model card. Parakeet runs on GPUs through NeMo and on CPUs and edge devices through sherpa-onnx (1.13.8, Apache-2.0), which makes it practical for edge AI and local apps.

How It Compares

Against Whisper, Parakeet trades breadth for precision. Whisper offers 99 languages, translation into English, the largest tooling ecosystem and an MIT license. Parakeet is faster, more accurate in English, returns native word timings and, by design, has less room to invent text over silence. A common pattern is Parakeet for English with Whisper as the multilingual fallback. Against hosted APIs such as Deepgram and AssemblyAI, Parakeet offers privacy, no per-hour bill and local control, at the cost of running it yourself. For speaker labels it pairs naturally with NVIDIA's TitaNet embeddings or pyannote (see speaker diarization). It shows how close open-weight models have come to commercial speech systems, although the best proprietary systems still lead the Open ASR Leaderboard's English average.