Audio & SpeechActive

Nemotron

NVIDIA's 600M-parameter English speech-to-text model, with a cache-aware streaming encoder and an RNN-T decoder. On Wiro it takes one recording and returns the transcript as plain text.

Speech to TextAudioGrid Landscape

Model ID

nemotron

Provider

nvidia

Updated

Oct 24, 2026, 15:20

Support Wiro on G2

wiro playground—nvidia/nemotron

Updated Oct 24, 2026, 15:20

Overview

Nemotron Speech Streaming En 0.6B is an English speech-to-text model from NVIDIA. The network has 600M parameters: a 24-layer cache-aware FastConformer encoder feeding an RNN-T (transducer) decoder.

NVIDIA designed it for live audio. The encoder keeps a cache of its self-attention and convolution states from one chunk to the next, so each new chunk is encoded once and earlier audio is never processed again. Buffered streaming setups re-run overlapping windows of audio instead. NVIDIA documents four chunk sizes between 0.08 and 1.12 seconds that can be switched at inference time without retraining, and in its tables the average word error rate drops as the chunk gets longer.

NVIDIA also positions the model for batch transcription, which is how it runs on Wiro: one recording goes in and the full transcript comes back as text.

What you can build

Searchable transcripts of recorded lectures and podcast episodes
Draft subtitle text for English video, to be aligned with timestamps in a later step
Voice notes converted to text before an LLM summarizes or routes them
A baseline for an English voice feature: run recordings from your own users through it and see how it handles their speech and your product's vocabulary

Inputs

One audio file, uploaded or passed as a URL. WAV, MP3, M4A, OGG, Opus and WebM are accepted. There is no language setting and no chunk-size control.

NVIDIA lists mono audio at 16 kHz as the model's native input and asks for at least 0.08 seconds of sound. Other sample rates are accepted here: one of the sample inputs on this page is a 24 kHz MP3.

Outputs

A plain-text English transcript, with no speaker labels or timestamps. Both sample transcripts on this page are lowercase and have no punctuation. NVIDIA lists punctuation and capitalization as built-in features of the model, so test a few of your own files before you build anything that depends on them. NVIDIA also notes that a file with no speech in it can return an empty transcript.

Limitations

English only. NVIDIA trained and evaluated the model on English speech.
Accuracy depends heavily on the material. At the 1.12 s chunk size, NVIDIA reports a word error rate of about 2.3% on LibriSpeech test-clean (read audiobook speech) and around 12% on AMI meeting recordings and Earnings22 earnings calls. Those scores are measured without punctuation and capitalization.

Safety & compliance

NVIDIA releases the model under the NVIDIA Open Model License and marks it ready for commercial and non-commercial use. It asks developers to make sure the model meets the requirements of their own industry and use case. Get consent from the people whose voices you transcribe.

API quick start

Run nemotron with a single API call.

POST https://api.wiro.ai/v1/Run/nvidia/nemotron

{
  "inputAudio": "https://your-cdn.com/input.mp3"
}

curl

curl -X POST "https://api.wiro.ai/v1/Run/nvidia/nemotron" \
  -H "Content-Type: application/json" \
  -H "x-api-key: YOUR_WIRO_API_KEY" \
  --data-binary @- <<'JSON'
{
  "inputAudio": "https://your-cdn.com/input.mp3"
}
JSON