Nvidia's open streaming speech model is live on DeepInfra at $0.00020 a minute

What Nvidia released

As deepinfra.com reported in its model listing, Nvidia's Nemotron Speech Streaming EN 0.6B is now available to rent by the minute. The listing title reads Nemo3-ASR-Streaming, but the page describes the model as a 600M-parameter system for English speech recognition. ASR stands for automatic speech recognition, the technical name for speech-to-text.

The key word is streaming. Instead of waiting for a recording to finish, the model writes text while the person is still talking. It adds punctuation and capital letters as it goes, so the output reads like a sentence rather than a raw stream of words.

Our earlier coverage said the listing was thin and left out price, license, size and languages. The full page fills in most of those gaps, and they are covered below.

How it keeps up with live speech

Older streaming systems often use a "buffered" approach. They cut audio into overlapping windows, so the computer analyzes some of the same sound twice. That costs processing power and adds delay.

Nvidia's design is called cache-aware FastConformer-RNNT. A cache is a short-term memory, and here it stores what the model already worked out about earlier audio. When a new slice arrives, the model processes only that slice and pulls the rest from memory. Nvidia says this saves computing effort and cuts delay without hurting accuracy.

Developers can choose how much audio the model handles per step: 80ms, 160ms, 560ms or 1120ms. Smaller chunks mean text appears sooner. DeepInfra runs the model at 560 ms lookahead, meaning it peeks that far ahead in the audio before committing to text.

What it costs and how to reach it

DeepInfra lists the price at $0.00020 per minute. It serves the model through a realtime WebSocket API, a kind of always-open connection that carries audio in and text out. The interface is compatible with OpenAI Realtime transcription, so code written for that style of service may need few changes. DeepInfra says it batches many simultaneous streams on each GPU, the chip that does the heavy math.

Developers who prefer other routes have several. The model page describes a lightweight C++ runtime for local use, Nvidia's NeMo toolkit, and the Hugging Face Transformers library from version 5.13.0. Nvidia also offers a hosted API on build.nvidia.com with a free key, and a public demo space on Hugging Face.

Training, license and limits

• Training data: The listing gives 530k hours of human-recorded audio. Most of it comes from Nvidia's Riva speech set (250k hours) plus the English portion of a dataset called Granary. Some transcripts were written by people and some by other speech models.
• License: The model is described as ready for commercial and non-commercial use, under the NVIDIA Open Model License Agreement. Read those terms before you ship a product.
• Language: English only, with audio that must be mono (one channel).
• Accuracy: The listing says the model was scored with word error rate, the share of words transcribed wrongly, on Hugging Face's open leaderboard datasets. The results tables are not in the text we reviewed, so we can't quote a score.

Newer versions already exist

The page carries dated notes. A March 12, 2026 update swapped in a checkpoint, meaning a saved set of trained model settings, trained on larger collections of speech. The older January 2026 checkpoint lives on a separate branch.

The newest note, dated June 4, 2026, points to Nemotron 3.5 ASR Streaming 0.6B. It stretches the same 600M-parameter size across 40 language-locales, which are language and region pairings. It also adds language prompts, optional automatic language detection and adjustable chunk sizes. If you need anything besides English, that is the version to look at.