Nemotron vs Parakeet: live words or finished files

Published October 23, 2026 · The Memory Notch team

Nemotron and Parakeet are NVIDIA speech-to-text models with different jobs. Parakeet transcribes a recording you already have. Nemotron transcribes audio while it is still arriving. On September 22, 2026, a separate company, Moondream, released Parakeet Redux and Parakeet Ultra: its own retrained versions of NVIDIA’s open Parakeet. Two days later, an open-source library ran them on a Mac. Like the original Parakeet, neither new version is a live model.

Checked September 29, 2026, on each model's Hugging Face page.

Memory Notch, our meeting recorder, runs the multilingual streaming version of Nemotron for its live transcript.

What changed this week

Moondream released Parakeet Redux and Parakeet Ultra on September 22, 2026. Both are built from Parakeet TDT v3, which we call v3 from here on, and keep its 25 languages and its licence. Redux shrinks the weights from 1.2 gigabytes to 178 megabytes so it runs on ordinary processors. Ultra keeps the full-size weights and makes fewer mistakes than the original on every published test.

Two days later, FluidAudio v0.17.3 added the pair. FluidAudio is an open-source library that runs speech models on a Mac. The release notes call them "two new batch models" and recommend the accurate one for new apps, while v3 stays the default.

So a desktop app can now pick a smaller or a more accurate version of the file model. It still gets a model that reads finished audio.

Nemotron vs Parakeet: same maker, different jobs?

Yes. The two come from NVIDIA at the same size, but one was trained to read finished audio and the other to read audio in small pieces as it arrives.

The file model has real strengths, and they come first. The model card reports low error rates on English and on recorded meetings, and the model gives a timestamp for every word. It reads up to 24 minutes in one pass on a data-centre graphics card. It has been out for over a year, and one port of it for laptops was downloaded about 284,000 times in the last 30 days.

The streaming model is newer and narrower in purpose. The maker rates 19 languages ready for transcription, including Hindi, Japanese and Arabic, which the file model doesn't cover. It adds punctuation and capital letters itself. It can also detect the spoken language and tag its output with it.

Which is more accurate?

On each vendor's own numbers, the new accurate version is the best of these models on finished recordings. No one has tested the streaming and file models on the same audio in the same way.

Word error rate is the share of words a model gets wrong, leaves out or adds. At 8%, about 8 words in every 100 are wrong, so lower is better.

Checked September 29, 2026.

Read the averages with care. The streaming model's average covers its strongest languages, and the file model's covers all of its own, so the two rows measure different sets. The streaming card lists no meeting test at all.

The fair reading is this. The file model has the better evidence on meetings and noise. The streaming model's numbers are close, and it scored them while streaming. It commits to its words with only 1.12 seconds of new audio at a time, before it hears how the sentence ends.

Why does live transcription need a streaming model?

A live transcript needs a model that reads each new piece of audio once and remembers what came before. Nemotron is trained to do exactly that. Running a file model live means re-reading overlapping slices of audio instead.

The maker calls this design cache-aware. The model keeps the working state of every layer from one chunk to the next, so a new chunk only adds work for the new audio. The card says this "eliminates redundant overlapping computations common in traditional 'buffered' streaming."

Buffered streaming is the usual way to run a file model live. You feed it a sliding window, and run it again each time the window moves.

Running the model again costs work. On one data-centre graphics card at 1.12-second chunks, the card reports about 2,400 live streams for the streaming model. A buffered file model twice its size manages about 400. That is a server figure, but a laptop pays the same kind of cost in battery and heat.

Running it again can also change words. When a model re-reads a window, the last few words can come out differently the second time. A streaming model commits each chunk's words once, so what you read during the call doesn't shift under you.

The file model's card points to a separate script for chunked streaming, and Moondream says its own engine accepts live audio. Those approaches work, but the model underneath was still trained to see more audio than it gets.

What are Parakeet Redux and Ultra?

Redux and Ultra are two retrained versions of v3 from Moondream, released on September 22, 2026. Redux is the small one and Ultra the accurate one.

Redux stores its encoder in 2-bit weights, which is how it gets to 178 megabytes. It loses the most in noise, where it got about 9 words in 100 wrong against under 7 for the original. On a laptop processor, the makers report it transcribing 38 seconds of audio per second.

The accurate version keeps the full-size weights and was trained further. It beat the original on all five published tests, including long talks and noisy audio.

The library's ports behave the same way on a laptop's machine-learning cores. On clean audiobook speech, the accurate version ran 126.7 times faster than real time, so an hour of audio takes about 28 seconds. The small one ran 83.9 times faster, and its first load compiles for several minutes.

Our comparison with Whisper sets the new pair against OpenAI's model on finished files.

Which runs on a Mac?

The two run on a Mac through ports built for its chips. The streaming model's own card lists only NVIDIA graphics cards on Linux.

For the file model, the choice is wide: the same library ships v3 and the two new versions.

For the streaming model, the library's makers publish a port of the multilingual version, which is the one our app runs. A meeting app still has to feed that engine the right audio at the right time. Four details decide how it behaves in a call.

Chunk size. Our app runs it at 1,120 ms chunks, the smallest size its library recommends when punctuation matters. Shorter chunks put words on screen sooner, with weaker punctuation.

Language, per chunk. The language is set to automatic, so it is detected for every chunk. A call that moves from English to Hindi and back is followed as it goes.

One model, two channels. The app transcribes your microphone and the call's audio separately, so each speaker stays on their own side. The two channels share one compiled copy of the model, which saves loading it twice.

No lost opening. The model can take several seconds to load, longer on the first launch after an update. The app holds audio that arrives before the model is ready rather than dropping it, which costs under 4 megabytes for a full minute of waiting. So the first word of the meeting gets transcribed like the rest.

Our guide to how AI transcription works walks through the rest of that pipeline, from the microphone to the speaker labels.

Which should you use for meetings, and which for files?

Use Nemotron when you want the words while the meeting is happening. Use Parakeet when you have a finished recording and want the best transcript of it.

For live meetings on a laptop, the streaming model is the right engine, for three reasons. It was built for streaming from the start. It covers more languages and can follow a speaker who switches between them. And it reads each second of audio once, which matters on a laptop that is running the call at the same time.

After the meeting, our app's enhanced transcript runs the same model again over the saved recording, in larger chunks. With no call to keep up with, it can take in more audio at each step.

What this comparison doesn't tell you

We have not benchmarked the two against each other ourselves. Every accuracy figure in this post is the vendor's, and most come from clean read speech rather than meetings.

No published test puts both models on the same audio in the same mode. Nemotron's card has no meeting test, so its number on a real call is unknown. The stream counts are for data-centre graphics cards, not a laptop.

The limits of our app are on the specifications page. A second model, Nemotron 3 Diarization, numbers the speakers rather than naming them, up to 8 per channel. Overlapping voices can be difficult to separate, and the app writes no automatic summaries.

Transcription runs on your computer, and the offline transcription guide explains what needs a connection. Running locally doesn't replace the consent of the people you record. You can try it free for recordings of up to 5 minutes.

Frequently asked questions

Sources

• Introducing Moondream Parakeet Redux and Parakeet Ultra, Moondream, September 22, 2026
• FluidAudio v0.17.3: Parakeet Ultra and Parakeet Redux, FluidInference on GitHub, September 24, 2026
• nvidia/nemotron-3.5-asr-streaming-0.6b model card, NVIDIA on Hugging Face, checked September 29, 2026
• nvidia/parakeet-tdt-0.6b-v3 model card, NVIDIA on Hugging Face, checked September 29, 2026
• Nemotron 3.5 ASR Streaming Multilingual CoreML port, FluidInference on Hugging Face, checked September 29, 2026
• Parakeet TDT v3 CoreML port, FluidInference on Hugging Face, checked September 29, 2026
• Memory Notch specifications

Related posts

• How does AI transcription work? Inside a Mac app
• Parakeet vs Whisper, and why we use neither live
• Silero VAD vs WebRTC VAD, and the end of speech
• Local AI note takers for Mac: what stays on it
• Offline transcription on a Mac
• Memory Notch specifications

Download Memory Notch · More from the blog