NVIDIA Nemotron 3.5 Transcribe API
nemotron-3.5-transcribeNemotron 3.5 Transcribe is NVIDIA's open-weight speech model, released on 28 October 2026. It turns recorded speech into text in 40 languages and detects the spoken language on its own; for 31 of those languages it can also translate to or from English, and it can label up to 8 speakers. On APIMODELS it is one model id, nemotron-3.5-transcribe (nemotron-3-5-transcribe, nemotron-transcribe and nvidia/nemotron-3-5-transcribe also work), on the audio task endpoint, and it is the first speech-to-text model here: the result is text, not an audio file. Send audio_url: a publicly downloadable MP3, WAV, M4A, AAC, FLAC, OGG, Opus or WebM file, an MP4 or MOV video with an audio track, or a data URI. One request takes up to 60 minutes and 200 MB. Set timestamps to true and the full result file adds a start time, an end time and a speaker label for every word (every character in Chinese). Set target_language to get a translation instead of a transcript; translation is only dependable when English is on one side, and it cannot be combined with timestamps (that request returns a 400). The call is asynchronous: you receive a taskId, poll GET /v1/audio/generations?task_id= or pass callback_url, and read text, duration and transcript_url from resultJson; the result file is kept for 30 days. In our own test, a 47-second English recording of two people, with timestamps on, came back in about 4 seconds with 143 words, each with start and end times and a speaker label. It suits meeting and interview minutes laid out by speaker, subtitles built from the word times, review of support and sales call recordings, searchable archives of lectures and earnings calls, turning multilingual recordings into English text, and producing transcripts to search or to hand to a language model; it does not suit live captions or realtime voice assistants. Limits to plan for: this is offline transcription of a whole file, not realtime streaming; speakers with similar voices can be merged into one; word end times are usually within tens of milliseconds, while 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; and not every task succeeds, for example when a link cannot be downloaded or a file has no decodable audio. Billing is $0.0025 per minute of audio ($0.15 per hour), prorated on the audio length rounded up to the whole second: 10 minutes is $0.025 and a full 60-minute file $0.15. 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 file whose length cannot be read returns a 400 with no charge, and failed tasks are not charged.
Transcribe audio to text
MP3, WAV, M4A, AAC, FLAC, OGG, Opus or WebM, or an MP4 / MOV video with an audio track. The API takes up to 60 minutes and 200 MB by URL; uploads here are limited to 10 MB for audio and 100 MB for video.
Adds a start time, an end time and a speaker label for every word to the full result JSON. No extra charge.
Returns a translation instead of a transcript. Dependable only when English is on one side: another language into English, or English into the language you pick. No extra charge.
$0.0025 per minute of audio ($0.15 per hour), prorated by the second. Timestamps, speaker labels and translation are included. Amounts below $0.0001 are billed $0.0001. Failed tasks are not charged.
Calling it from code? See the API docs.
Transcript
The transcript will appear here as text
Speech to Text in 40 Languages
Transcribes recorded speech and detects the language on its own, from English, Mandarin, Japanese and Korean to Arabic, Hindi, Thai and Vietnamese
Word Timestamps and Speakers
With timestamps on, the result file carries a start time, an end time and a speaker label for every word, for up to 8 speakers
Translation to or from English
Set target_language to get a translation instead of a transcript: 31 languages into English, or English into them
Up to 60 Minutes per File
MP3, WAV, M4A, AAC, FLAC, OGG, Opus or WebM, or an MP4 / MOV video with an audio track, up to 200 MB
Only Success Is Charged
Failed tasks are not charged, and a file whose length cannot be read is rejected with a 400 before anything is billed
$0.0025 per Minute of Audio
Billed by the second: 10 minutes $0.025, 60 minutes $0.15. Timestamps, speaker labels and translation cost nothing extra
Nemotron 3.5 Transcribe API: speech to text in 40 languages, with timestamps and speakers
Nemotron 3.5 Transcribe is NVIDIA's open-weight speech model, released on 28 October 2026: it takes a recorded audio or video file of up to 60 minutes and returns the transcript as text, in 40 languages with automatic language detection.
On APIMODELS the Nemotron 3.5 Transcribe API is a single model id, nemotron-3.5-transcribe, on POST /v1/audio/generations. A failed Nemotron 3.5 Transcribe task is not charged; a finished one costs $0.0025 per minute of audio, with word timestamps, speaker labels and translation included.
- Model ID
nemotron-3.5-transcribe- Input
- Audio or video, up to 60 min / 200 MB
- Languages
- 40, detected automatically
- Output
- Text + full result JSON
- Billing
- $0.0025 per minute of audio
What Nemotron 3.5 Transcribe does
A speech-to-text model for recorded files: the transcript, who said it, and when.
Speech to text in 40 languages
Language detected automatically
Nemotron 3.5 Transcribe transcribes English, Mandarin Chinese, Japanese, Korean, the major European languages, Arabic, Hindi, Thai, Vietnamese and more, 40 in all. You do not tell it which language the recording is in.
audio_urlWord-level timestamps
A start and end time for every word
With timestamps on, the full result file lists every word with its start and end time, which is what subtitles, search-in-audio and clip cutting are built from. Chinese comes back character by character.
timestampsSpeaker labels
Up to 8 speakers
The same switch adds a speaker label to every word and a list of who spoke when, so a meeting or an interview reads as a dialogue. Speakers with similar voices can be merged into one.
Translation to or from English
31 languages paired with English
Set a target language and Nemotron 3.5 Transcribe returns a translation instead of a transcript. It is dependable when English is on one side; it cannot be combined with timestamps.
target_languageAudio or video in
Up to 60 minutes, 200 MB
MP3, WAV, M4A, AAC, FLAC, OGG, Opus and WebM audio, or an MP4 or MOV video with an audio track, sent as a public link or a data URI. There is no need to extract the audio first.
Text plus a full result file
text, duration, transcript_url
The task result carries the full text and the audio length; transcript_url points to a JSON file with the words, speakers and segments. The file is kept for 30 days.
What Nemotron 3.5 Transcribe is a good fit for
Six jobs it suits, each with what you send, which switch you set and what comes back, plus the one it does not.
Meeting and interview minutes
Send the recording of a call, a meeting or an interview with timestamps: true. You get the full text, plus a result file in which every word carries a speaker label and diarization lists who spoke when, so the transcript can be laid out speaker by speaker.
timestampsSubtitles for video and podcasts
Send the MP4 or MOV itself, or the episode audio, with timestamps: true. Build SRT or VTT cues yourself from the start and end times in words. Start times can drift by about half a second, so check where each cue begins.
timestampswordsSupport calls and sales recordings
Send long, multi-speaker call recordings with timestamps: true. Speaker labels tell apart up to 8 people, so agent and customer turns can be separated for review, keyword checks and analysis. Voices that sound alike can be merged, so spot-check the labels.
Archives of lectures, courses and earnings calls
Send long recordings one file per request, up to 60 minutes and 200 MB each; cut anything longer into parts. The text makes every recording searchable, at $0.15 per hour of audio.
Multilingual content to or from English
Set target_language to en and a talk in Japanese, German or Spanish comes back as English text; or send English audio and pick one of the 31 other languages. The text field holds the translation. Timestamps are not available on translation requests.
target_languageTranscripts for search and LLM summaries
Send voice notes, recordings or media files without extra options and store the returned text. It can be indexed for search, or passed as plain text to one of the language models on apimodels.app, which the same API key and balance also cover.
Live captions and realtime voice assistants
Nemotron 3.5 Transcribe works offline on a whole recorded file and returns nothing until the task finishes, so it cannot caption a live stream or drive a voice assistant. For a live two-way voice conversation, use Grok Voice 3.0.
How the Nemotron 3.5 Transcribe API works
One required field, two optional switches, and limits that are checked before a task is created.
Transcribe
Send the file to POST /v1/audio/generations with model nemotron-3.5-transcribe. Add timestamps: true when you need word times and speakers.
audio_urlRequired. A publicly downloadable audio or video file, or a data URI; up to 60 minutes and 200 MB.timestampsOptional boolean, false by default. True adds word-level times and speaker labels to the full result file.callback_urlOptional. We POST the finished task here; leave it out and poll the GET endpoint.
Translate
Same endpoint and file field; add target_language and the text comes back in that language. Leave timestamps off.
audio_urlRequired, same rules as above.target_languageOne of 32 codes such as en, zh, ja, ko, fr, de, es. Dependable for another language into English, or English into one of these.
| Nemotron 3.5 Transcribe input | Accepted | If it is out of range |
|---|---|---|
| File format | MP3, WAV, M4A, AAC, FLAC, OGG, Opus, WebM; MP4 or MOV with an audio track | Rejected or the task fails; nothing is charged |
| Length | Up to 60 minutes per request | 400 before a task is created |
| File size | Up to 200 MB | Rejected or the task fails; nothing is charged |
| Readable length | We must be able to measure the audio | 400 before a task is created, no charge |
| timestamps + target_language | One or the other | 400 before a task is created |
| Output | text, duration and transcript_url (full result JSON, kept 30 days) | — |
There is no source-language, prompt or output-format parameter. For SRT or VTT subtitles, request timestamps: true and build the cues from the word times in the result file.
Nemotron 3.5 Transcribe pricing: $0.0025 per minute of audio
One rate, counted on the audio length and prorated by the second.
$0.15
One hour of audio
60 minutes × $0.0025. Word timestamps, speaker labels and translation do not change the price, and failed Nemotron 3.5 Transcribe tasks are not charged.
| Audio length | Nemotron 3.5 Transcribe price |
|---|---|
| 47 s (our test clip) | about $0.002 |
| 1 minute | $0.0025 |
| 10 minutes | $0.025 |
| 30 minutes | $0.075 |
| 60 minutes (maximum per request) | $0.15 |
2-minute voice memo
$0.005
2 min × $0.0025
20-minute podcast segment with speakers
$0.05
20 min × $0.0025, timestamps included
40-minute lecture translated into English
$0.10
40 min × $0.0025, translation included
3-hour workshop, split into three files
$0.45
3 requests × 60 min × $0.0025
How the amount is counted: we read the audio length when the task is created, round it up to the whole second and multiply by the per-minute rate. Amounts are in units of $0.0001, so anything smaller is billed $0.0001 (audio of about 2 seconds or less). If the length cannot be read, the Nemotron 3.5 Transcribe request returns a 400 and nothing is charged.
What we measured with Nemotron 3.5 Transcribe
One run of our own. It is one example, not a promised speed.
Input
47 s of English, two speakers
A short two-person conversation, sent with timestamps: true.
Output
143 words with times and speakers
Every word came back with a start time, an end time and a speaker label, alongside the full text.
Time and price
About 4 seconds, about $0.002
Longer files take longer. Poll every few seconds, or pass callback_url and wait for the result.
Limits of Nemotron 3.5 Transcribe to plan for
What the model does not do, and where its output needs a second look.
Offline only, no realtime streaming
It transcribes a whole recorded file after you submit it. There are no partial results while it runs and no live captioning.
60 minutes and 200 MB per request
Longer recordings have to be split and submitted as separate requests.
Timestamps or translation, not both
Word times and speaker labels exist only for transcripts. A request with timestamps: true and target_language returns a 400.
Similar voices can be merged
Up to 8 speakers are told apart, but two people who sound alike can end up under one label.
Timestamps are approximate
Word end times are usually within tens of milliseconds. Start times can be off by about half a second, more noticeably in English.
Runs are not identical
Transcribing the same audio twice can change a few words. Keep the result you got rather than assuming a rerun will match it.
Chinese names and numbers
In Chinese output, foreign names are often transliterated and some numbers are written as Chinese characters.
Non-English language pairs
Translation is dependable to or from English. Going directly between two other languages is not guaranteed.
How to call the Nemotron 3.5 Transcribe API
From API key to transcript in four steps.
Get an API key
Sign up at apimodels.app and create a key in the console. One key and one USD balance cover Nemotron 3.5 Transcribe and every other model.
Host the recording
Put the audio or video file at a link anyone can download directly. Links behind a login, and temporary links that expire within minutes, make the task fail.
Send the request
POST /v1/audio/generations with model nemotron-3.5-transcribe and audio_url, plus timestamps or target_language if you need them. The response holds a taskId.
Read the transcript
Poll GET /v1/audio/generations?task_id= until the state is completed, or wait for the callback. JSON.parse resultJson once, read text, and fetch transcript_url for the word-level file within 30 days.
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
}'
# { "code": 200, "msg": "success", "data": { "taskId": "clxxx", "state": "pending", "model": "nemotron-3.5-transcribe" } }
curl "https://api.apimodels.app/v1/audio/generations?task_id=TASK_ID" -H "Authorization: Bearer YOUR_API_KEY"
# data.resultJson (a JSON string): { "text": "...", "duration": 47.2, "transcript_url": "https://r2.apimodels.app/transcripts/clxxx.json" }Why run Nemotron 3.5 Transcribe on apimodels.app
What you can check for yourself, before the price.
Failed tasks are not charged
If a Nemotron 3.5 Transcribe task ends in failed, the charge is returned automatically and the reason is in failMsg.
Bad inputs stop before billing
A missing audio_url, an unreadable file length, a file over 60 minutes, or timestamps combined with translation all return a clear 400 before a task exists.
Async tasks with a callback
Create, then poll the task id or receive the result at callback_url. An hour-long file never holds a connection open.
One API key for every model
The same key and balance call Nemotron 3.5 Transcribe, LLMs, text-to-speech, image and video models, so transcribe-then-summarize needs one integration.
Your own copy of the result file
The full result JSON is stored at transcript_url for 30 days, so you can fetch the word-level data again without re-running the task.
Per-minute pricing, by the second
$0.0025 per minute of audio, measured from your file. Pay as you go, no subscription; the playground shows the estimate before you run it.
Nemotron 3.5 Transcribe vs Whisper-large-v3: word error rate
These are the numbers NVIDIA published in the Nemotron 3.5 Transcribe Technical Report (28 October 2026); they are not our own measurements. The metric is word error rate (WER), the share of words a model gets wrong, so lower is better. NVIDIA takes the baseline figures from the Hugging Face Open ASR Leaderboard.
In the report's English, long-form and multilingual tables, Nemotron 3.5 Transcribe has the lowest average word error rate among the open-weight models listed. It is not first overall: the same report lists commercial systems that score better on long-form and multilingual audio.
| Benchmark | What it covers | Nemotron 3.5 Transcribe | Whisper-large-v3 |
|---|---|---|---|
| Open ASR (average of 9 English test sets) | General English speech | 4.71% | 6.66% |
| Earnings21/22 and CORAAL (average) | Long-form audio | 10.98% | 13.92% |
| AMI | Meeting recordings | 7.81% | 13.63% |
| Average of 7 languages | Multilingual speech | 4.62% | 7.94% |
Source: NVIDIA Nemotron 3.5 Transcribe Technical Report, Table 6 (English), Table 7 (long-form) and Table 9 (multilingual). The 7 languages are German, French, Italian, Spanish, Portuguese, Dutch and Hindi.
For recordings with several speakers, Table 14 of the same report gives cpWER (a word error rate that also counts words assigned to the wrong speaker, lower is better) on whole meeting recordings, macro-averaged over AMI, CALLHOME, Mixer 6 and NOTSOFAR-1: Nemotron 3.5 Transcribe 15.80%, against 16.83% for DiCoW v3.3 + DiariZen-v2.
Sources for Nemotron 3.5 Transcribe
Model facts and benchmark numbers on this page come from NVIDIA's own publications. Prices, limits and the 47-second test are ours.
- NVIDIA Nemotron 3.5 Transcribe Technical Report (28 October 2026) ↗
Word error rates in Tables 6, 7 and 9, and the multi-speaker cpWER in Table 14.
- Nemotron 3.5 Transcribe model card on Hugging Face ↗
Open weights, supported languages and stated limitations.
- NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt (28 October 2026) ↗
NVIDIA's release article on Hugging Face.
- Hugging Face Open ASR Leaderboard ↗
Where the baseline figures in the comparison come from.
About Nemotron 3.5 Transcribe
Nemotron 3.5 Transcribe is a Speech-to-Text API provided by NVIDIA. Nemotron 3.5 Transcribe is NVIDIA's open-weight speech model, released on 28 October 2026. It turns recorded speech into text in 40 languages and detects the spoken language on its own; for 31 of those languages it can also translate to or from English, and it can label up to 8 speakers. On APIMODELS it is one model id, nemotron-3.5-transcribe (nemotron-3-5-transcribe, nemotron-transcribe and nvidia/nemotron-3-5-transcribe also work), on the audio task endpoint, and it is the first speech-to-text model here: the result is text, not an audio file. Send audio_url: a publicly downloadable MP3, WAV, M4A, AAC, FLAC, OGG, Opus or WebM file, an MP4 or MOV video with an audio track, or a data URI. One request takes up to 60 minutes and 200 MB. Set timestamps to true and the full result file adds a start time, an end time and a speaker label for every word (every character in Chinese). Set target_language to get a translation instead of a transcript; translation is only dependable when English is on one side, and it cannot be combined with timestamps (that request returns a 400). The call is asynchronous: you receive a taskId, poll GET /v1/audio/generations?task_id= or pass callback_url, and read text, duration and transcript_url from resultJson; the result file is kept for 30 days. In our own test, a 47-second English recording of two people, with timestamps on, came back in about 4 seconds with 143 words, each with start and end times and a speaker label. It suits meeting and interview minutes laid out by speaker, subtitles built from the word times, review of support and sales call recordings, searchable archives of lectures and earnings calls, turning multilingual recordings into English text, and producing transcripts to search or to hand to a language model; it does not suit live captions or realtime voice assistants. Limits to plan for: this is offline transcription of a whole file, not realtime streaming; speakers with similar voices can be merged into one; word end times are usually within tens of milliseconds, while 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; and not every task succeeds, for example when a link cannot be downloaded or a file has no decodable audio. Billing is $0.0025 per minute of audio ($0.15 per hour), prorated on the audio length rounded up to the whole second: 10 minutes is $0.025 and a full 60-minute file $0.15. 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 file whose length cannot be read returns a 400 with no charge, and failed tasks are not charged. Through APIMODELS platform, you can access this model via a unified API with transparent pay-as-you-go pricing. Current pricing: per minute of audio (billed by the second): $0.0025, 10 minutes: $0.025, 30 minutes: $0.075, 60 minutes (maximum per request): $0.15.
Related model: Grok Voice 3.0 — when you need a live two-way voice conversation instead of transcribing a recorded file — realtime speech-to-speech over WebSocket, $0.104/min
Related model: AI Dubbing — when you want the recording re-voiced in another language rather than a text translation
Related model: Voice Isolator — when the recording is noisy — isolate the voice first, then transcribe the cleaned file
Key Features
- Speech to Text in 40 Languages -- Transcribes recorded speech and detects the language on its own, from English, Mandarin, Japanese and Korean to Arabic, Hindi, Thai and Vietnamese
- Word Timestamps and Speakers -- With timestamps on, the result file carries a start time, an end time and a speaker label for every word, for up to 8 speakers
- Translation to or from English -- Set target_language to get a translation instead of a transcript: 31 languages into English, or English into them
- Up to 60 Minutes per File -- MP3, WAV, M4A, AAC, FLAC, OGG, Opus or WebM, or an MP4 / MOV video with an audio track, up to 200 MB
- Only Success Is Charged -- Failed tasks are not charged, and a file whose length cannot be read is rejected with a 400 before anything is billed
- $0.0025 per Minute of Audio -- Billed by the second: 10 minutes $0.025, 60 minutes $0.15. Timestamps, speaker labels and translation cost nothing extra
Use Cases
Voiceover & Narration
Generate professional-grade voiceovers for videos, animations, and ads with diverse voice options.
Podcast Production
Quickly produce podcast audio content with support for multi-character dialogue.
Audiobook Creation
Convert text content into natural, fluid speech for audiobook production.
Multilingual Dubbing
AI-powered multilingual dubbing and translation to help content reach global audiences.
Why APIMODELS
- Unified API -- One API key to access all models, no need to register on multiple platforms
- Transparent Pricing -- Pay-as-you-go with per-call pricing published on every model page, ideal for indie developers and startups
- Instant Access -- Start right after signup; pay by card, PayPal, Alipay or USDT
- Full Documentation -- Detailed API docs with code examples in cURL, Python, and Node.js
Frequently Asked Questions
What is the difference between APIMODELS and the official NVIDIA API?
APIMODELS offers the same Nemotron 3.5 Transcribe model through our aggregation platform. We provide a unified API interface so you do not need separate accounts for each provider - one API key to access all models.
What is the Nemotron 3.5 Transcribe API?
Nemotron 3.5 Transcribe is NVIDIA's open-weight speech model, released on 28 October 2026, and the Nemotron 3.5 Transcribe API turns a recorded audio or video file into text. It transcribes 40 languages and detects the language on its own, can add word-level timestamps with speaker labels for up to 8 speakers, and can translate between English and 31 other languages. On apimodels.app it is one model id, nemotron-3.5-transcribe, on POST /v1/audio/generations, billed at $0.0025 per minute of audio; failed tasks are not charged.
Which languages does Nemotron 3.5 Transcribe support?
Forty for transcription, 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 and Vietnamese. Thirty-one of them can also be translated to or from English. You do not pass a source-language parameter.
Which audio and video files does Nemotron 3.5 Transcribe accept?
MP3, WAV, M4A, AAC, FLAC, OGG, Opus and WebM audio, or an MP4 or MOV video with an audio track. Pass it as audio_url: a link that anyone can download directly, or a data URI. One request takes up to 60 minutes and 200 MB; split longer recordings and submit the parts separately. A file whose length cannot be read is rejected with a 400 before anything is charged.
How do I get word-level timestamps and speaker labels?
Send timestamps: true. The task result still carries the full text, and the result file at transcript_url (also resultUrls[0]) adds words, a list of {word, start, end, speaker}, plus diarization segments and text_with_speakers. Up to 8 speakers are told apart; Chinese is returned character by character. In our test, a 47-second English recording of two people came back with 143 words, each with start and end times and a speaker label. Timestamps and speaker labels cost nothing extra.
Can Nemotron 3.5 Transcribe translate speech?
Yes. Set target_language to one of the 32 supported codes (en, zh, ja, ko, fr, de, es and so on) and the text comes back translated instead of transcribed. Quality is only dependable when English is on one side: another language into English, or English into one of those languages. Direct translation between two non-English languages is not guaranteed. Translation cannot be combined with timestamps: true; that request returns a 400. It is included in the per-minute price.
Does Nemotron 3.5 Transcribe support realtime or streaming transcription?
No. It transcribes a whole recorded file offline: you create a task, then poll it or wait for the callback. In our test a 47-second clip with timestamps finished in about 4 seconds, but that is one measurement, not a latency promise, and partial results are not streamed while the file is processed. For a live two-way voice conversation, Grok Voice 3.0 on apimodels.app is the realtime option.
How accurate is Nemotron 3.5 Transcribe, and what are its known limits?
It is built for clean recorded speech, and a few limits are worth planning for. Speakers with similar voices can be merged into one. Word end times are usually within tens of milliseconds, while 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. If the result has truncated: true, the output hit its ceiling, usually because of repetition, and should be discarded or reviewed; silent: true means almost no speech was found.
How does Nemotron 3.5 Transcribe compare with Whisper large-v3?
On the word error rates NVIDIA published in its Technical Report of 28 October 2026 (lower is better), Nemotron 3.5 Transcribe is ahead of Whisper-large-v3 on all four comparisons: 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 (German, French, Italian, Spanish, Portuguese, Dutch, Hindi). These are NVIDIA's figures, with baselines taken from the Hugging Face Open ASR Leaderboard, not measurements by apimodels.app. In those three tables it has the lowest average error among open-weight models; the same report lists commercial systems that do better on long-form and multilingual audio.
NVIDIA Nemotron 3.5 Transcribe Technical Report (Tables 6, 7, 9) ↗How much does the Nemotron 3.5 Transcribe API cost?
$0.0025 per minute of audio, which is $0.15 per hour, prorated on the audio length rounded up to the whole second. A 47-second clip is about $0.002, 10 minutes $0.025, 30 minutes $0.075 and a 60-minute file $0.15. Word timestamps, speaker labels and translation are included at no extra charge. Amounts are in units of $0.0001, so anything smaller is billed $0.0001 (audio of about 2 seconds or less). Failed tasks, and files whose length cannot be read, are not charged.
How do I call the Nemotron 3.5 Transcribe API and read the result?
Create an account at apimodels.app and generate an API key; one key and one USD balance cover every model. Then send: 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}'. The response holds a taskId. Poll GET /v1/audio/generations?task_id= until state is completed, then JSON.parse resultJson and read text, duration and transcript_url. The result file is kept for 30 days.
Can I try Nemotron 3.5 Transcribe for free?
Eligible new accounts get $0.10 of trial credit, which at $0.0025 per minute covers about 40 minutes of audio. After that, billing is pay-as-you-go with no subscription, and top-ups start at $10. The playground on this page lets you upload a file or paste a link, shows the estimated price from the audio's length before you run it, and returns the transcript as text you can copy.
What can Nemotron 3.5 Transcribe do?
Nemotron 3.5 Transcribe supports: Speech to Text, 40 Languages, Auto-Detected, Word-Level Timestamps, Speaker Labels (up to 8), Translation to/from English, Audio or Video, up to 60 min, $0.0025 per Minute. See the APIMODELS docs for full parameters and call examples.
Can I use the Nemotron 3.5 Transcribe API from any country, including mainland China?
Yes. APIMODELS does not restrict access by country: the site and the API answer requests from anywhere, and apimodels.app and api.apimodels.app are reachable directly from mainland China without a proxy. You call Nemotron 3.5 Transcribe with one APIMODELS key, so you do not need a separate account with the model's developer.
What payment methods are supported? Can I pay with PayPal or crypto?
Yes to both. You can top up with a card (Visa, Mastercard and others, through Stripe), PayPal, Alipay, or USDT on TRON (TRC20) or BNB Smart Chain (BEP20). Card, PayPal and Alipay top-ups are credited immediately; USDT is credited once the transfer confirms on-chain. Auto top-up is optional, and a $50 top-up adds a $1 bonus ($3 on a $100 top-up). The balance is in USD and billed per use, with no subscription.