Live transcription needs a different kind of model. Whisper and Parakeet read a block of audio and then produce text, which works well for files but adds a delay when someone is talking right now. A streaming model processes audio in small chunks and keeps its state between them, so words appear a fraction of a second after they are spoken. NVIDIA’s Nemotron Speech family is built for exactly this.
The short version
- What it is: 600M-parameter streaming speech recognition models from NVIDIA, released in 2026.
- Languages: the multilingual model card lists 32 locales that transcribe out of the box. TranscribeX offers live transcription in 30+ languages.
- Output: punctuation and capitalization while you watch, which the lighter Parakeet EOU live model lacks.
- In TranscribeX: the model behind live captions, the live meeting transcript and live speaker labels. Part of Pro.
The Nemotron ASR models
| nemotron-speech-streaming-en-0.6b | nemotron-3.5-asr-streaming-0.6b | |
|---|---|---|
| Released | March 2026 (first checkpoint January 2026) | June 2026 |
| Parameters | 600M | 600M |
| Architecture | FastConformer, cache-aware RNNT | FastConformer, cache-aware RNNT with language prompt |
| Languages | English | 40 locales: 19 transcription-ready, 13 broad coverage, 8 need fine-tuning |
| Chunk sizes | 80, 160, 560, 1,120 ms | 80, 160, 320, 560, 1,120 ms |
| Punctuation and capitals | ✓ | ✓ |
| Language detection | n/a | ✓Optional |
| Published WER | Open ASR avg: 6.93% at 1.12 s, 7.67% at 160 ms, 8.43% at 80 ms | FLEURS avg (transcription-ready locales): 8.84% at 1.12 s, 10.38% at 80 ms |
| Licence | NVIDIA Open Model License | OpenMDW-1.1 |
From the NVIDIA model cards for nemotron-speech-streaming-en-0.6b and nemotron-3.5-asr-streaming-0.6b. The two WER figures use different test sets and are not comparable with each other. Japanese, Korean and Chinese are scored by character error rate on the multilingual card. Checked October 2026.
The chunk size is the main dial. Smaller chunks mean text appears sooner but the model has less context, so it makes more mistakes: on the English model, average WER goes from 6.93% with 1.12-second chunks to 8.43% with 80 ms chunks. For captions you read along with, a delay of around half a second to a second is usually comfortable. For voice control, the shortest chunks matter more.
The multilingual model’s transcription-ready locales are US and UK English, US and European Spanish, French and Canadian French, Italian, Brazilian and European Portuguese, Dutch, German, Turkish, Russian, Arabic, Hindi, Japanese, Korean, Vietnamese and Ukrainian. NVIDIA lists a further 13 languages as broad coverage, including Polish, Swedish, Czech, Danish, Finnish, Mandarin Chinese and Romanian.
What Nemotron does in TranscribeX
TranscribeX Pro uses Nemotron for its live features:
- Live captions. A floating caption window that transcribes whatever your Mac is playing, such as a video, a webinar or a call, in real time.
- The live meeting transcript. During a detected meeting, a window shows the transcript as people talk, with an AI chat panel to ask what you missed and real-time translation under each line.
- Live speaker diarization. Speaker labels appear in the live transcript as people talk, not only after the call.
When the meeting ends, TranscribeX transcribes the saved recording with your chosen file model, such as Parakeet or Whisper Large V3, for the final transcript, summary and exports. Streaming models trade a little accuracy for speed, so the after-call pass is usually the cleaner text.
Nemotron vs the other live options
| Nemotron | Parakeet EOU 160ms | Apple On-Device | |
|---|---|---|---|
| Languages | 30+ | English | 30 locales |
| Punctuation | ✓ | ✗ | ?Not stated by Apple |
| Size | 600M params | 120M params | Built into macOS |
| TranscribeX tier | Pro | Pro | Free |
From the NVIDIA model cards and Apple’s WWDC25 session 277. Apple’s locale count is from the TranscribeX model list.
The Parakeet EOU model is smaller and very quick but English only and unpunctuated. Apple On-Device streams too and costs nothing. Nemotron is the one to use when you need punctuated live text in many languages.
Running a streaming model locally
Cloud caption services transcribe on their own servers, so the audio is uploaded while you speak. With Nemotron in TranscribeX, the audio is transcribed on your Mac and nothing is uploaded. That also means it keeps working on a plane or a bad connection. AI chat can use the model built into the app as well; only online AI providers you choose to connect need the internet. More in offline transcription on a Mac.
“TranscribeX has been great! I am on 20+ calls weekly and this tool has become invaluable for call transcripts, and without any subscription or sending my data off to the cloud! I've emailed support on a few occasions and Ethan has been extremely responsive, top notch support!”
Nemotron ASR FAQ
What is Nemotron ASR?
Nemotron ASR, or Nemotron Speech, is a family of streaming speech recognition models from NVIDIA. They use a cache-aware FastConformer design that transcribes audio in small chunks, from 80 ms up to about a second, with punctuation and capitalization.
How many languages does Nemotron support?
NVIDIA’s multilingual nemotron-3.5-asr-streaming-0.6b lists 40 language-locales, of which 32 transcribe out of the box. TranscribeX offers live transcription in 30+ languages.
Is Nemotron the same as Parakeet?
No. Both come from NVIDIA and share the FastConformer encoder, but Parakeet TDT models are built for files, while Nemotron is built for streaming. TranscribeX uses Parakeet for file transcription and Nemotron for live captions.
Does Nemotron need an NVIDIA GPU?
Not in TranscribeX. The model runs locally on your Mac after a one-time download.