Nemotron 3.5 Transcribe API reference

Pricing

Amount = audio length (rounded up to the whole second) ÷ 60 × $0.0025, with the length measured by us when the task is created. Timestamps, speaker labels and translation cost nothing extra. Amounts are in units of $0.0001, so anything smaller is billed $0.0001 (audio of about 2 seconds or less). A request whose audio length cannot be read returns a 400 with no charge, and failed tasks are not charged.

Authentication

All requests carry the API key in the header:

Authorization: Bearer YOUR_API_KEY

Create a task and poll it

# Nemotron 3.5 Transcribe — speech to text. audio_url is the only required input.
# timestamps: true adds word-level times and speaker labels to the full result file.
# Billed $0.0025 per minute of audio, prorated by the second.
curl -X POST https://api.apimodels.app/v1/audio/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nemotron-3.5-transcribe",
    "audio_url": "https://example.com/meeting.mp3",
    "timestamps": true
  }'

# Response (async — a task is created):
# { "code": 200, "msg": "success", "data": { "taskId": "clxxx", "state": "pending",
#     "model": "nemotron-3.5-transcribe", "reservedCredits": 0.025, "currency": "USD" } }

# Poll every few seconds until completed
curl "https://api.apimodels.app/v1/audio/generations?task_id=TASK_ID" \
  -H "Authorization: Bearer YOUR_API_KEY"
# { "code": 200, "data": { "state": "completed",
#     "resultUrls": ["https://r2.apimodels.app/transcripts/clxxx.json"],
#     "resultJson": "{\"text\":\"Thanks for joining...\",\"duration\":600.2,\"transcript_url\":\"https://r2.apimodels.app/transcripts/clxxx.json\",\"task\":\"transcribe\"}" } }

A task moves through pending → processing → completed or failed. This is offline transcription of a whole file: there is no realtime streaming and no partial result while it runs. In our own test, a 47-second English recording of two people (timestamps on) came back in about 4 seconds with 143 words, each with start and end times and a speaker label; that is one measurement, not a promise, and longer files take longer. On failed, the reason is in data.failMsg and the charge is returned.

Parameters

There is no source-language parameter: the language is detected automatically. There is also no prompt, vocabulary or output-format parameter; for SRT or VTT subtitles, request timestamps: true and build the cues from the word times.

Result fields

When the task completes, data.resultJson in the query result and the callback is a JSON string; JSON.parse it once to get the shape below. text is the full transcript (the translation when target_language is set); transcript_url points to the full result JSON, and the same address is in data.resultUrls[0].

// JSON.parse(data.resultJson) on a completed task
{
  "resultUrls": ["https://r2.apimodels.app/transcripts/clxxx.json"],
  "text": "Thanks for joining. Let's start with the roadmap...",   // full transcript, or the translation
  "duration": 47.2,                                                // seconds of audio
  "transcript_url": "https://r2.apimodels.app/transcripts/clxxx.json",
  "task": "transcribe"                                             // or "translate"
  // "silent": true      almost no speech was found; text is empty
  // "truncated": true   the output hit its ceiling, usually repetition; discard or review
}

The full result JSON (transcript_url) looks like this. The last three keys exist only with timestamps: true; for Chinese, words holds one entry per character. Like other result files it is kept for 30 days, so download or re-host it promptly.

// The file at transcript_url (kept 30 days)
{
  "task": "transcribe",
  "duration": 47.2,
  "text": "Thanks for joining. Let's start with the roadmap...",
  // The three keys below are present only with timestamps: true
  "text_with_speakers": "<spk:0> Thanks for joining. <spk:1> Let's start with the roadmap...",
  "words": [
    { "word": "Thanks", "start": 0.32, "end": 0.61, "speaker": "0" },
    { "word": "for", "start": 0.61, "end": 0.74, "speaker": "0" }
  ],
  "diarization": [
    { "speaker": 0, "start": 0.32, "end": 1.83 }
  ]
}

Callback (callback_url)

To skip polling, pass callback_url in the create request. When the task reaches completed or failed we send an HTTP POST (JSON) to it. Failed deliveries are retried with backoff for up to 30 minutes, until your endpoint returns a 2xx. The URL must be publicly reachable; localhost and private addresses return a 400 at create time. Polling keeps working either way.

POST {your callback_url}
Content-Type: application/json

{
  "code": 200,
  "msg": "success",
  "data": {
    "taskId": "clxxx...",
    "model": "nemotron-3.5-transcribe",
    "state": "completed",                 // or "failed"
    "param": "<JSON string>",             // your request, JSON.parse once
    "resultJson": "<JSON string> | null", // {"text": "...", "duration": 47.2, "transcript_url": "https://...json"}, JSON.parse once
    "failCode": null,
    "failMsg": null
  }
}

Supported languages

• Transcription (40, detected automatically) — English, Mandarin Chinese, Japanese, Korean, French, German, Spanish, Portuguese, Italian, Russian, Dutch, Polish, Czech, Slovak, Hindi, Bulgarian, Croatian, Danish, Estonian, Finnish, Hungarian, Latvian, Lithuanian, Romanian, Slovenian, Swedish, Thai, Turkish, Ukrainian, Arabic, Bengali, Greek, Hebrew, Indonesian, Maltese, Marathi, Tagalog, Tamil, Telugu, Vietnamese.
• Translation (target_language, 32 codes) — en zh ja ko fr de es pt it ru nl pl cs sk hi bg hr da et fi hu lv lt ro sl sv th uk ar el he mt. Only translation to or from English is dependable; direct translation between two non-English languages is not guaranteed.

Limits

• Offline transcription of a whole file; realtime streaming is not supported.
• Up to 60 minutes and 200 MB per request; split longer recordings and submit the parts separately.
• Timestamps (timestamps: true) and translation (target_language) cannot be combined; that request returns a 400.
• Speaker labels tell apart up to 8 speakers; speakers with similar voices can be merged into one.
• Word end times are usually within tens of milliseconds; start times can be off by about half a second, more noticeably in English.
• Running the same audio twice can change a few words.
• In Chinese output, foreign names are often transliterated and some numbers are written as Chinese characters.
• silent: true means almost no speech was found and text is empty; truncated: true means the output hit its ceiling (usually repetition) and should be discarded or reviewed.

Errors

Parameter problems come back synchronously as a 400 at creation ({"code": 400, "msg": "..."}): no task is created and nothing is charged. The msg strings below are the current wording (N is the actual number).

Failures during processing end in state: failed with the reason in data.failMsg, and the charge is returned automatically. The common ones:

• 401: the API key is missing or invalid.
• 402: the balance does not cover this audio.

Notes

• audio_url must be an address anyone can download directly; links behind a login and temporary links that expire within minutes make the task fail.
• The result file (transcript_url) is deleted after 30 days, so download or re-host it promptly.
• Only upload recordings you have the right to process.

Try it in the Playground. Use cases, worked price examples and FAQs are on the Nemotron 3.5 Transcribe model page.

Sources

Model facts (release date, supported languages, speaker count) and benchmark numbers come from NVIDIA's own publications. On the word error rates in NVIDIA's Technical Report (WER, lower is better; not measured by us), Nemotron 3.5 Transcribe against Whisper-large-v3: 4.71% vs 6.66% on the Open ASR average of 9 English test sets, 10.98% vs 13.92% on long-form audio (Earnings21/22 and CORAAL average), 7.81% vs 13.63% on AMI meeting recordings, and 4.62% vs 7.94% averaged over 7 languages.

• NVIDIA Nemotron 3.5 Transcribe Technical Report (28 October 2026; word error rates in Tables 6, 7 and 9)
• Nemotron 3.5 Transcribe model card on Hugging Face
• NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt (28 October 2026)
• Hugging Face Open ASR Leaderboard (source of the baseline figures)

Related