Parakeet
Updated
Definition
Parakeet is a family of open-source speech recognition models from NVIDIA, built on a transducer architecture, which can emit words as audio arrives. That makes it a common engine inside live dictation tools, where Whisper is built to read a whole chunk first.
Parakeet models are trained for the same job as Whisper but with a different shape: instead of reading a whole chunk before writing, a transducer produces tokens continuously, so text can appear while you are still speaking. On Apple silicon the smaller Parakeet variants run in real time with room to spare, which makes them a natural fit for dictation where the words should land at key-up.
The trade is coverage. Parakeet models are trained on a smaller set of languages than Whisper, so tools that use one for streaming pair it with a batch model for everything else.
Transducer versus encoder-decoder
An encoder-decoder model like Whisper is a reader and a writer: the encoder digests a window of audio, then the decoder writes the transcript. A transducer joins the two so that every new slice of audio can produce a new token without waiting for the window to close. That is what streaming needs, and it is also why transducers tend to be more stable on long audio: they never get to "reconsider" a whole passage and repeat it.
Where it fits in a dictation tool
Parakeet is the engine you want running while the key is held. It gives the live transcript in the notch or the pill, confirms you were heard, and leaves almost nothing to do at key-up. The final text quality still depends on the formatting stage that follows, and on a dictionary for names, exactly as with any ASR model.
Parakeet v2 vs v3: which version does what
The Parakeet models most people mean today are the two TDT 0.6B releases. Both have 600 million parameters, both add punctuation and capitalization, both return timestamps, and both are released under the CC BY 4.0 license.
The difference is language. Version 2 is English only and was trained on about 120,000 hours of English speech. Version 3 extends it to 25 European languages and detects which one is being spoken without being told. NVIDIA reports an average word error rate on the Hugging Face Open ASR Leaderboard of 6.05% for v2 and 6.34% for v3.
- Western and northern Europe: Danish, Dutch, English, Finnish, French, German, Swedish.
- Southern Europe: Greek, Italian, Maltese, Portuguese, Spanish.
- Central Europe: Croatian, Czech, Hungarian, Polish, Slovak, Slovenian.
- Baltic: Estonian, Latvian, Lithuanian.
- Eastern Europe: Bulgarian, Romanian, Russian, Ukrainian.
How to run Parakeet on your own audio
NVIDIA publishes Parakeet through its NeMo toolkit for Python. The model card lists Linux as the supported operating system and says the model is optimized for NVIDIA graphics cards, so treat a run on a Mac as an experiment. It needs at least 2 GB of memory to load, and more memory lets it take longer audio.
- Install a recent version of PyTorch, then install the toolkit with: pip install -U nemo_toolkit['asr']
- Convert your recording to the format the model expects, 16 kHz mono in wav or flac. With ffmpeg: ffmpeg -i input.m4a -ar 16000 -ac 1 audio.wav
- Start Python and run: import nemo.collections.asr as nemo_asr
- Load the model with: asr_model = nemo_asr.models.ASRModel.from_pretrained(model_name="nvidia/parakeet-tdt-0.6b-v3")
- Transcribe with: output = asr_model.transcribe(['audio.wav'])
- Print the text with: print(output[0].text)
- Add timestamps=True to the transcribe call if you want word and segment timestamps.
Parakeet vs Whisper, by the published facts
Both are open speech recognition models you can download, and both write punctuated text. They differ in size, coverage, license, and what they were designed to do. NVIDIA describes Parakeet TDT 0.6B v3 as designed for high-throughput transcription, and its model card also documents a streaming mode that processes audio in chunks of a few seconds.
- Languages: Parakeet TDT 0.6B v3 covers 25 European languages. Whisper was trained on English plus 98 other languages.
- Size: Parakeet TDT 0.6B has 600 million parameters. Whisper ranges from 39 million (tiny) to 1550 million (large).
- License: Parakeet is CC BY 4.0, which requires attribution. Whisper is MIT.
- Translation: Whisper can translate other languages into English. The Parakeet TDT cards describe transcription only.
- Long audio: Parakeet handles up to 24 minutes in one pass, or up to 3 hours with local attention. Whisper moves a 30-second window across the file.
- Input: Parakeet expects 16 kHz mono wav or flac. Whisper uses ffmpeg to read common audio formats.
Questions people ask.
What is Parakeet in speech recognition?
An open-source family of ASR models from NVIDIA built to stream: words come out as audio goes in. It is often paired with Whisper, which covers more languages but works in batch.
Is Parakeet better than Whisper?
For live dictation in a supported language, it feels better because text appears as you speak. For language coverage and for transcribing files, Whisper is broader. Good tools use each for what it is best at.
What languages does Parakeet support?
Parakeet TDT 0.6B v2 supports English only. Version 3 supports 25 European languages, including English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, and Ukrainian, and detects the language automatically.
Is Parakeet free for commercial use?
Yes. NVIDIA releases Parakeet TDT 0.6B under the CC BY 4.0 license and states that the model is ready for commercial and non-commercial use. The license requires attribution.
How long an audio file can Parakeet transcribe?
NVIDIA says Parakeet TDT 0.6B v3 handles audio up to 24 minutes long with full attention on an A100 80GB graphics card, or up to 3 hours with local attention.
