NVIDIA Nemotron 3.5 Transcribe
NVIDIA Nemotron 3.5 TranscribeA speech-understanding model that brings multilingual transcription, speaker attribution, translation, and audio understanding into one workflow for recorded audio.
Model details
About NVIDIA Nemotron 3.5 Transcribe
NVIDIA Nemotron 3.5 Transcribe helps turn recorded conversations into transcripts and useful insights. It combines multilingual speech recognition, speaker attribution, speech-to-text translation, contextual prompting, and audio questions in a single open-weight model.
Use plain-language prompts to provide relevant product names and domain terminology, guide transcription style, or ask questions about a recording. Bringing these tasks into one model reduces the need to coordinate separate components for each step of a speech-understanding workflow.
Nemotron 3.5 Transcribe is built for recorded-audio applications, including post-call analytics, meeting tools, media transcription, and voice-of-customer analysis.
Key capabilities
Multilingual transcription: Transcribe recordings in supported languages, including multilingual conversations with code-switching.
Speaker attribution: Produce speaker-aware transcripts that help preserve who said what.
Speech-to-text translation: Generate translated text from recorded speech.
Contextual prompting: Supply relevant names, terminology, and task instructions to guide recognition.
Audio understanding: Ask questions about a recording to support review and analysis.
Open-weight deployment: Retain control over deployment and integrate speech understanding into your application.
Using the model on Baseten
Nemotron 3.5 Transcribe processes recorded audio rather than live audio streams. Provide an audio recording and a text prompt describing the desired task, such as verbatim transcription, clean transcription, or translation.
Review generated transcripts, speaker labels, and inferred insights against your application’s requirements before using them in downstream workflows.
1import base64
2import os
3
4from openai import OpenAI
5
6client = OpenAI(
7 api_key=os.environ["BASETEN_API_KEY"],
8 base_url=(
9 "https://model-<model_id>.api.baseten.co"
10 "/environments/production/sync/v1"
11 ),
12)
13
14with open("audio.wav", "rb") as audio_file:
15 audio = base64.b64encode(audio_file.read()).decode("utf-8")
16
17response = client.chat.completions.create(
18 model="nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B",
19 messages=[
20 {
21 "role": "user",
22 "content": [
23 {
24 "type": "text",
25 "text": (
26 "Audio transcription. Language: as spoken. "
27 "Wording: verbatim, including fillers and repetitions. "
28 "Formatting: punctuation and capitalization. "
29 "Output: transcript only."
30 ),
31 },
32 {
33 "type": "audio_url",
34 "audio_url": {
35 "url": f"data:audio/wav;base64,{audio}"
36 },
37 },
38 ],
39 }
40 ],
41 temperature=0,
42 max_tokens=1024,
43 extra_body={
44 "chat_template_kwargs": {"enable_thinking": False}
45 },
46)
47
48print(response.choices[0].message.content)
