01
Transcription in 40 Languages
NVIDIA organizes its 40 transcription languages into three tiers: 13 Leading, 15 Production-ready, and 12 Supported. The model can detect the spoken language automatically.
Input
Click to upload or drag and drop
Supported formats: MP4, QUICKTIME, WEBM Maximum file size: 200MB; Supported formats: MPEG, WAV, X-WAV, AAC, MP4, OGG, FLAC, OPUS Maximum file size: 200MB
Audio to transcribe: a publicly downloadable URL. Common formats work (MP3, WAV, M4A/AAC, FLAC, OGG/Opus, WEBM), and so do videos with an audio track (MP4, MOV); only the sound is used. Up to 60 minutes and 200 MB per file. The language is detected automatically.
Also return word-level start and end times and speaker labels in the result file. Transcription only: not available together with target_language. The example output was produced with timestamps on and lists only the first 40 words; the API result file contains every word.
Leave empty (default) to transcribe in the spoken language. Pick a language to translate the speech into it instead. Translation is guaranteed between English and the other listed languages: non-English speech into English (en), or English speech into another language.
Output
output typetext
Loading JSON viewer...
README
Complete guide to using nvidia/nemotron-3-5-transcribe
Nemotron 3.5 Transcribe API for Multilingual Speech to Text
Transcribe audio across 40 languages with NVIDIA's Nemotron 3.5 Transcribe model. On Kie.ai, a single request can return the transcript, word-level timestamps, speaker segments for up to 8 speakers, and speech translation between English and 31 languages, billed per second of audio.
New to Nemotron 3.5 Transcribe? Read what it is and how it compares →
New to Nemotron 3.5 Transcribe? Read what it is and how it compares →
What the Nemotron 3.5 Transcribe API Can Do
02
Word Timestamps and Speaker Labels in One Pass
Get start and end times for every word and speaker segments for up to 8 speakers in the same result. NVIDIA supports speaker diarization in 13 languages.
03
Speech Translation to and from English
Alongside transcription, the model translates speech between English and 31 languages, so one call can turn a recording in a supported language into English text, or English speech into text in a supported language.
04
Built for Long-Form Audio
NVIDIA trained the model on audio files up to one hour long and reports results on long-form test sets such as earnings calls.
Nemotron 3.5 Transcribe Accuracy (NVIDIA-Reported)
NVIDIA reports the following word error rates (WER, lower is better) across public test sets. These figures come from NVIDIA's published results and were not measured on Kie.ai.
| Test set | What it covers | Nemotron 3.5 Transcribe | Whisper-large-v3 |
|---|---|---|---|
| Open ASR (9 English sets, average) | 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% |
| 7-language average | Multilingual speech | 4.62% | 7.94% |
How to Use the Nemotron 3.5 Transcribe API on Kie.ai
- 1
Step 1. Get Your API Key
Create an account on Kie.ai, open the API Key Management page, and generate an API key. Pass it on every request in the Authorization: Bearer YOUR_API_KEY header. - 2
Step 2. Create a Transcription Task
Send a POST request to /api/v1/jobs/createTask with the Nemotron 3.5 Transcribe model and your audio file URL in audio_url. Set timestamps to get word-level timing, and set target_language when you want the speech translated; speaker labels are included in the result. - 3
Step 3. Read the Result
Transcription runs as an asynchronous task. Once the job finishes, read the result to retrieve the transcript, speaker segments and, if you requested them, word timestamps.
Tips Before You Ship
- 01
Check for Silent Audio First
NVIDIA notes that silent or non-speech input can generate text that was never spoken. Skip silent recordings and trim long stretches of silence before submitting audio. - 02
Know the Weaker Cases
According to NVIDIA, code-like terms, Dutch audio, and speech that switches between languages, such as mixed Chinese and English, are transcribed less accurately than with Whisper-large-v3. - 03
Plan for Whole-File Processing
Because the model transcribes complete files without streaming partial results or returning confidence scores, design your workflow around a finished transcript.
Why Use the Nemotron 3.5 Transcribe API on Kie.ai
Same Key Across the Kie.ai Market
All models in the Kie.ai market use the same authentication method: an Authorization: Bearer YOUR_API_KEY header with the key from your Kie.ai API Key Management page.
Free Credits to Start
New Kie.ai users receive 80 free credits for testing. Your account has one credit balance that all generation services draw from, and you can check it with the Get Remaining Credits API.
Nothing to Host or Install
Kie.ai provides this model as an API with a playground; there is no app to download or install. There are no GPUs or model weights to manage.
Typical Uses for the Nemotron 3.5 Transcribe API
Because one request can return text, timestamps, and speaker labels, the model suits workloads that need more than a plain transcript.
Meeting and Interview Notes
Convert long meetings and interviews into structured transcripts with each speaker's lines separated and timed.
Subtitles and Captions
Line up captions with video using word-level start and end timestamps.
Multilingual Content
Transcribe source recordings across 40 languages, or translate speech between English and 31 languages in the same call.
Earnings Calls and Long Recordings
Turn earnings calls and other long recordings into searchable text.
Frequently Asked Questions
How much does the Nemotron 3.5 Transcribe API cost?
Usage is billed per second of audio, and word timestamps, speaker labels, and translation are included at no extra cost. See the pricing card above for the current rate.
Which languages does Nemotron 3.5 Transcribe support?
The model supports 40 transcription languages across three NVIDIA tiers, along with speaker diarization in 13 languages, speech translation between English and 31 languages, and automatic language detection.
Does the Nemotron 3.5 Transcribe API return timestamps and speaker labels?
Yes. One request can return the transcript together with word-level start and end times and speaker segments for up to 8 speakers. NVIDIA supports speaker diarization in 13 languages.
How accurate is Nemotron 3.5 Transcribe?
Across 9 public English test sets, NVIDIA reports an average word error rate of 4.71%, compared with 6.66% for Whisper-large-v3. Refer to the accuracy table above for additional benchmarks.
What audio formats and lengths can I send?
Pass an audio URL in a common format such as WAV, FLAC, MP3, OGG, or M4A. Audio is processed as 16 kHz mono, and NVIDIA trained the model on files up to one hour long.
Does Nemotron 3.5 Transcribe support real-time streaming?
No. The model transcribes complete audio files offline rather than streaming partial results.
Is the Nemotron 3.5 Transcribe API free?
No. Every request uses credits at the rate shown in the pricing card above, though new Kie.ai users receive 80 free credits for testing.
How do I get a Nemotron 3.5 Transcribe API key?
Sign up on Kie.ai and generate a key on the API Key Management page. That same key works across every model in the Kie.ai market.