NVIDIA · Speaker diarization

NVIDIA Nemotron 3 Diarization

Find who spoke when with up to eight speakers, overlapping speech and precise speaker turns. Returns timestamps and speaker labels without a transcript.

nemotron-3-diarization

$0.010000 / audio minuteTry in Console

Current model-specific price from our catalog. Requests require your project API key and sufficient balance. Availability is checked at execution time.

Input and capabilities

Upload a short audio or video file using multipart/form-data: at most 30 seconds and 20 MiB. Use WAV for the examples below. Container compatibility depends on the model.

diarization
segment timestamps

Identify up to eight speakers, including overlapping speech. This endpoint returns who spoke when; it does not produce a transcript or recognize a speaker’s identity. Labels belong to the current recording. Silence is a valid result with an empty speakers array.

No language or expected speaker count is required. Presets control how the model processes the recording internally; every preset uses a file upload and a complete response, not a realtime connection.

API endpoint

POST https://api.speechinfra.com/v1/inference/nemotron-3-diarization

Authenticate with your Speechinfra project key. Read the quickstart and billing guide.

The first request may wait while the model starts. Examples allow up to 16 minutes, including time to receive an error response; later requests can reuse the running worker. Avoid resubmitting while a request is pending.

cURL

curl --fail-with-body --max-time 960 "https://api.speechinfra.com/v1/inference/nemotron-3-diarization" \
  -H "Authorization: Bearer $SPEECH_API_KEY" \
  -F "file=@clip.wav" \
  --form-string "preset=offline" \
  --form-string "onset=0.5" \
  --form-string "offset=0.5" \
  --form-string "pad_onset=0" \
  --form-string "pad_offset=0" \
  --form-string "min_duration_on=0" \
  --form-string "min_duration_off=0" \
  --form-string "return_rttm=false"

Python

# pip install httpx
import os
import httpx

headers = {"Authorization": "Bearer " + os.environ["SPEECH_API_KEY"]}
with open("clip.wav", "rb") as audio:
    response = httpx.post(
        "https://api.speechinfra.com/v1/inference/nemotron-3-diarization",
        headers=headers, files={"file": ("clip.wav", audio, "audio/wav")},
        data={
    "preset": "offline",
    "onset": "0.5",
    "offset": "0.5",
    "pad_onset": "0",
    "pad_offset": "0",
    "min_duration_on": "0",
    "min_duration_off": "0",
    "return_rttm": "false"
}, timeout=960,
    )
response.raise_for_status()
print(response.json())

Node.js / fetch

// Node.js 22+, no SDK required.
import { readFile } from 'node:fs/promises';
const form = new FormData();
form.set('file', new Blob([await readFile('clip.wav')]), 'clip.wav');
for (const [key, value] of Object.entries({
  "preset": "offline",
  "onset": "0.5",
  "offset": "0.5",
  "pad_onset": "0",
  "pad_offset": "0",
  "min_duration_on": "0",
  "min_duration_off": "0",
  "return_rttm": "false"
})) form.set(key, value);
const response = await fetch("https://api.speechinfra.com/v1/inference/nemotron-3-diarization", {
  method: 'POST', body: form, signal: AbortSignal.timeout(960000),
  headers: { Authorization: 'Bearer ' + process.env.SPEECH_API_KEY },
});
if (!response.ok) throw new Error(await response.text());
console.log(await response.json());

Request parameters

Generated from the API schema for this deployment. Conditional parameters apply only when their controlling setting is enabled.

Parameter	Type	Details
`file` Required	string	Short audio or video file. Direct API limit: 30 seconds, 20 MiB.
`preset`	string	Internal chunking preset. The endpoint always returns the complete recording; this is not a realtime stream. enum: offline, low_latency, very_low_latency, ultra_low_latency · default: "offline" All accepted values [ "offline", "low_latency", "very_low_latency", "ultra_low_latency" ]
`onset`	number	Speech start probability threshold. default: 0.5 · min: 0 · max: 1
`offset`	number	Speech end threshold; must not exceed onset. default: 0.5 · min: 0 · max: 1
`pad_onset`	number	Seconds added before a speech segment. default: 0 · min: 0 · max: 1
`pad_offset`	number	Seconds added after a speech segment. default: 0 · min: 0 · max: 1
`min_duration_on`	number	Minimum retained speech segment duration in seconds. default: 0 · min: 0 · max: 10
`min_duration_off`	number	Fill speech gaps shorter than this many seconds. default: 0 · min: 0 · max: 10
`return_rttm`	boolean	Include standard RTTM speaker turns alongside JSON. default: false

Complete request schema

{
  "type": "object",
  "additionalProperties": false,
  "required": [
    "file"
  ],
  "properties": {
    "file": {
      "type": "string",
      "format": "binary",
      "description": "Short audio or video file. Direct API limit: 30 seconds, 20 MiB."
    },
    "preset": {
      "type": "string",
      "enum": [
        "offline",
        "low_latency",
        "very_low_latency",
        "ultra_low_latency"
      ],
      "default": "offline",
      "description": "Internal chunking preset. The endpoint always returns the complete recording; this is not a realtime stream."
    },
    "onset": {
      "type": "number",
      "minimum": 0,
      "maximum": 1,
      "default": 0.5,
      "description": "Speech start probability threshold."
    },
    "offset": {
      "type": "number",
      "minimum": 0,
      "maximum": 1,
      "default": 0.5,
      "description": "Speech end threshold; must not exceed onset."
    },
    "pad_onset": {
      "type": "number",
      "minimum": 0,
      "maximum": 1,
      "default": 0,
      "description": "Seconds added before a speech segment."
    },
    "pad_offset": {
      "type": "number",
      "minimum": 0,
      "maximum": 1,
      "default": 0,
      "description": "Seconds added after a speech segment."
    },
    "min_duration_on": {
      "type": "number",
      "minimum": 0,
      "maximum": 10,
      "default": 0,
      "description": "Minimum retained speech segment duration in seconds."
    },
    "min_duration_off": {
      "type": "number",
      "minimum": 0,
      "maximum": 10,
      "default": 0,
      "description": "Fill speech gaps shorter than this many seconds."
    },
    "return_rttm": {
      "type": "boolean",
      "default": false,
      "description": "Include standard RTTM speaker turns alongside JSON."
    }
  }
}

Response schema

duration, start and end are in seconds. speakers contains speaker turns, so one speaker may appear many times. Turns can overlap. Set return_rttm=true to also receive RTTM annotations.

Example speaker turns

{
  "duration": 2.5,
  "speakers": [
    {
      "speaker": "speaker_0",
      "start": 0.2,
      "end": 1.8,
      "overlap": true
    },
    {
      "speaker": "speaker_1",
      "start": 1.5,
      "end": 2.3,
      "overlap": true
    }
  ],
  "model": "nemotron-3-diarization",
  "provider": "runpod"
}

View complete response schema

{
  "$defs": {
    "DiarizationTurn": {
      "properties": {
        "speaker": {
          "title": "Speaker",
          "type": "string"
        },
        "start": {
          "title": "Start",
          "type": "number"
        },
        "end": {
          "title": "End",
          "type": "number"
        },
        "overlap": {
          "default": false,
          "title": "Overlap",
          "type": "boolean"
        }
      },
      "required": [
        "speaker",
        "start",
        "end"
      ],
      "title": "DiarizationTurn",
      "type": "object"
    }
  },
  "properties": {
    "duration": {
      "title": "Duration",
      "type": "number"
    },
    "speakers": {
      "items": {
        "$ref": "#/$defs/DiarizationTurn"
      },
      "title": "Speakers",
      "type": "array"
    },
    "model": {
      "title": "Model",
      "type": "string"
    },
    "provider": {
      "title": "Provider",
      "type": "string"
    },
    "rttm": {
      "anyOf": [
        {
          "type": "string"
        },
        {
          "type": "null"
        }
      ],
      "default": null,
      "title": "Rttm"
    }
  },
  "required": [
    "duration",
    "speakers",
    "model",
    "provider"
  ],
  "title": "DiarizationResult",
  "type": "object"
}