Workbench
Output
Ready for a new runRelated models and next tasks
Missing the model you need?
NVIDIA · OPEN-WEIGHTS SPEECH MODEL
NVIDIA Nemotron 3.5 Transcribe: hour-long recordings to text in 40 languages, with word timings, speaker turns or a translation.
Key numbers
- Languages
- 40
- Nemotron 3.5 Transcribe detects the spoken language itself; there is nothing to select.
- Longest file
- 60 min
- One recording of up to an hour per request; longer files are rejected free of charge.
- Timestamps
- Per word
- Start, end and speaker number for each word when word timestamps are on.
- Translation pairs
- 31
- Speech translated to or from English instead of transcribed.
Before you spend
All modes
- One recording per request, from 1 second to 60 minutes. Longer files are rejected before they run and cost nothing. Uploads are capped at 90 MB; public links can point to larger files.
- Word timestamps and translation cannot be used together. There is no language setting, custom vocabulary or streaming output on Nemotron 3.5 Transcribe; the result arrives when the whole file is done.
- Similar-sounding voices can be merged into one speaker, and word start times can drift by up to about half a second. Audio that loops the same clip over and over can come back cut short; in our hour-long loop test the text stopped after a few repetitions.
What it does

Language detection
Forty languages, no language box to fill in
Upload the file and the language is worked out from the voice itself, so an archive of mixed-language recordings needs no sorting first. In our tests a Japanese weather report came back in Japanese script, and an English call kept its punctuation and question marks.
Words and speakers
A start, an end and a speaker for every word
With word timestamps on, each word in the result carries its start and end time in seconds and a speaker number counting up from 0. Our two-voice radio call split into speaker 0 and speaker 1 at the right turns. Start times can drift by up to about half a second and similar voices may merge, which suits caption lines but not frame-accurate sync.
Translation
Translate in the same request
Set target_language and the result is the speech rendered in that language instead of a transcript. In our runs, a Spanish narration came back in English and an English call came back in Japanese. Translation works to or from English with 31 other languages, but it cannot be combined with word timestamps.

Long recordings
A 58-minute file back in about 52 seconds
One request takes a recording of up to 60 minutes, and the result arrives when the whole file is done. In our tests a 58-minute audiobook chapter finished in about 52 seconds, and the same file with word timestamps produced 10,594 timed words. Anything over 60 minutes is rejected before it runs and costs nothing.
Where it fits
Podcast and interview transcripts
Speaker-separated text for quotes, show notes and search.
Caption timing
Word-level timings for subtitle files in the spoken language.
Multilingual archives
Recordings in any of 40 languages transcribed without tagging them first.
Rough translations
A first English draft of a foreign-language recording, or English speech rendered for another market.
Model reference
Nemotron 3.5 Transcribe API reference
Nemotron 3.5 Transcribe API specifications
- Model
- NVIDIA Nemotron 3.5 Transcribe (open weights, OpenMDW-1.1)
- Endpoint
- Speech to text
- Languages
- 40 for transcription, detected automatically
- Translation
- To or from English, 31 other languages
- Input
- MP3, WAV, MP4 or WebM upload up to 90 MB, or a public link; 1 second to 60 minutes
- Output
- Full text; with word timestamps, each word's start, end and speaker number
- Speed
- A 58-minute file in about 52 seconds in our tests
How to call the Nemotron 3.5 Transcribe API
Upload or link the recording
MP3, WAV, MP4 or WebM up to 90 MB as an upload, or a public link to a longer file; anything up to 60 minutes of speech.
Choose plain text or timings
Leave word timestamps off for a fast plain transcript; turn them on for captions and speaker turns.
Or translate instead
Pick a target language to get the speech in that language. Nemotron 3.5 Transcribe translates to or from English.
Read the result
The transcript arrives as text, and with timestamps also as a list of words with times and speaker numbers.
The short version
Nemotron 3.5 Transcribe is NVIDIA's open-weights answer to long, multilingual recordings: send up to an hour of speech in any of 40 languages and get the full text back in about a minute, with word timings and speaker turns when you need captions, or a translation when you need it in another language.
Nemotron 3.5 Transcribe API questions
All Nemotron models2
Audio models1
Nemotron 3.5 TranscribeCurrentSpeech to text$0.00003per second of input audio
