Models / NVIDIA
Audio

NVIDIA Nemotron 3 ASR Streaming 0.6B

Low-latency English streaming transcription with configurable chunk sizes down to 80ms

Try now
read docs

About model

Nemotron 3 ASR Streaming 0.6B is NVIDIA's 600-million-parameter English speech recognition model, engineered to deliver high-quality transcription across both low-latency streaming and high-throughput batch workloads. Built on a cache-aware FastConformer-RNNT architecture, it processes only new audio chunks while reusing cached encoder context, removing the redundant overlapping computation of traditional buffered streaming. Transcripts arrive with native punctuation and capitalization, and configurable chunk sizes of 80ms, 160ms, 560ms, and 1120ms let teams choose their operating point on the latency-accuracy curve at inference time, with no retraining required. Available on Together AI.

‍

Lowest Chunk Size

80ms

Low-latency streaming for real-time voice agents

Runtime Chunk Sizes

4

80ms to 1120ms, adjustable at inference with no retraining

Punctuation & Capitalization

Native

Production-ready transcripts straight from streaming audio

Model key capabilities
  • Cache-Aware Streaming: Processes only new audio chunks while reusing cached encoder context for efficient continuous transcription
  • Runtime Flexibility: Configurable 80ms, 160ms, 560ms, and 1120ms chunk sizes to balance latency and accuracy at inference time
  • Formatted Transcripts: Native punctuation and capitalization on every transcription, ready for production use
  • Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure

  • API usage

    • cURL
    • Python
    • Typescript

    Endpoint:

    nvidia/nemotron-3-asr-streaming-0.6b

    curl -X POST "https://api.together.xyz/v1/audio/transcriptions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -F "model=nvidia/nemotron-3-asr-streaming-0.6b" \
  -F "language=en" \
  -F "response_format=json" \
  -F "timestamp_granularities=segment"

    from together import Together

client = Together()
response = client.audio.transcribe(
    model="nvidia/nemotron-3-asr-streaming-0.6b",
    language="en",
    response_format="json",
    timestamp_granularities="segment"
)
print(response.text)

    import Together from "together-ai";

const together = new Together();

const response = await together.audio.transcriptions.create(
  model: "nvidia/nemotron-3-asr-streaming-0.6b",
  language: "en",
  response_format: "json",
  timestamp_granularities: "segment"
});
console.log(response)

  • Model card

    Architecture Overview:
    • Cache-aware FastConformer encoder with an RNNT decoder, 600 million parameters
    • Maintains caches for encoder self-attention and convolution layers, so each processed frame is strictly non-overlapping
    • Designed for continuous audio streams in low-latency voice applications as well as high-throughput batch transcription

    Training Methodology:
    • Trained primarily on NVIDIA's Riva ASR training set (250,000 hours) and the English portion of the Granary dataset
    • Hybrid labeling combining human transcriptions with model-generated transcripts

    Performance Characteristics:
    • Word error rate measured on the HuggingFace Open ASR Leaderboard datasets
    • Accuracy improves as chunk size grows while remaining competitive at the lowest-latency 80ms setting

    ‍

  • Prompting

    Together AI API Access:
    • Access Nemotron 3 ASR Streaming 0.6B via Together AI APIs using the endpoint nvidia/nemotron-3-asr-streaming-0.6b
    • Authenticate using your Together AI API key in request headers
    • Send or stream 16kHz monochannel audio and receive English transcriptions with punctuation and capitalization
    • Select a chunk size (80ms, 160ms, 560ms, or 1120ms) to match your latency target
    • Available on Together AI serverless and dedicated infrastructure

    ‍

  • Applications & use cases

    Real-Time Voice Agents:
    • Run the listening layer for conversational agents with chunk sizes tuned as low as 80ms
    • Keep transcription context across a continuous stream without buffered recomputation
    • Pass formatted transcripts straight into LLM pipelines on Together AI

    Live Captioning & Accessibility:
    • Caption live events, streams, and meetings in English with low end-to-end delay
    • Deliver readable output with punctuation and capitalization applied automatically
    • Tune chunk size per event to balance caption speed against accuracy

    High-Volume Batch Transcription:
    • Transcribe recorded call libraries and media archives on the same endpoint used for streaming
    • Use larger chunk sizes for maximum accuracy on offline workloads
    • Standardize one English ASR model across live and batch pipelines

    ‍

Related models
new
Transcribe
NVIDIA Parakeet TDT 0.6B V3 Realtime

  • Price/min

    $0.0035
  • Parameters
    600M
No items found.
Audio
new
Transcribe
NVIDIA Parakeet TDT 0.6B v3

  • Price/min

    $0.0015
No items found.
Transcribe
new
Transcribe
NVIDIA Nemotron 3.5 ASR

  • Price/min

    $0.0045/min
  • Parameters
    0.6B
No items found.
Transcribe
Model specifications
Model data
  • Model provider
    NVIDIA
  • Type
    Audio
  • Deployment
    Serverless
    On-Demand Dedicated
  • Endpoint
    nvidia/nemotron-3-asr-streaming-0.6b
  • Parameters
    600M
  • Price

    $0.0015 / min per minute of audio

  • Input modalities
    Audio
  • Output modalities
    Text
  • Category
    Transcribe
Run in Playground
Quickstart docs
Deploy model