Nemotron 3.5 ASR — NVIDIA Streaming Speech to Text
Try Nemotron AI audio-to-text live in your browser
Nemotron is NVIDIA's family of open models, and Nemotron 3.5 ASR is its 0.6B multilingual streaming speech-recognition model. Upload a clip or paste an audio URL below and get a punctuated transcript in seconds.
Use auto-detect unless you know the main spoken language.
Higher acceleration uses smaller streaming chunks for lower latency; none keeps 1.12s chunks for the best accuracy, while full uses 0.08s chunks for the lowest latency.
to run the demo. New accounts get free starter credits; each run then costs 1 credit.
Heads up: this demo sends audio to NVIDIA's hosted Nemotron 3.5 ASR endpoint on fal through our server. If the provider is unavailable, your audio is safe to transcribe in the Whisper AI workspace.
This is an independent guide. Nemotron and Nemotron 3.5 ASR are products of NVIDIA; the demo runs through fal's hosted nvidia/nemotron-asr-multilingual/asr endpoint.
What is Nemotron 3.5 ASR?
Nemotron 3.5 ASR is a 600-million-parameter, multilingual streaming speech-recognition model from NVIDIA. It is the speech member of the Nemotron family of open models — released with open weights, training data and recipes, and available as NVIDIA NIM microservices.
Under the hood it uses a Cache-Aware FastConformer-RNNT architecture with language-ID prompt conditioning, so one checkpoint handles roughly 40 language-locales and emits native punctuation and capitalization.
The interactive panel at the top of this page lets you try Nemotron AI speech-to-text right now. It calls the hosted fal endpoint from our server so the API key stays private while the open weights remain available on Hugging Face for self-hosting.
Why Nemotron 3.5 ASR stands out
Streaming recognition, a compact multilingual checkpoint and runtime-tunable latency make Nemotron a practical speech-to-text model.
Streaming, low-latency transcription
Nemotron 3.5 ASR is built for real-time streaming recognition. NVIDIA reports sub-100ms time-to-final on short chunks — treat the figure as a vendor benchmark, but the design targets live captioning and voice agents.
Cache-Aware FastConformer-RNNT
The model uses a Cache-Aware FastConformer encoder with an RNNT decoder, so it reuses streaming context efficiently instead of re-processing buffered audio on every step.
Multilingual in one checkpoint
A single 0.6B checkpoint covers roughly 40 language-locales (about 36 languages) with language-ID prompt conditioning — no per-language model swapping required.
Native punctuation & capitalization
Transcripts come back with punctuation and capitalization baked in, so the output reads naturally without a separate restoration pass.
Runtime-configurable latency
Switch between chunk sizes — roughly 1.12s, 0.56s, 0.32s and 0.08s — to trade accuracy for speed at request time via the acceleration setting, no redeploy needed.
Compact 0.6B footprint
At 600M parameters the model is small enough to serve at high concurrency. NVIDIA cites markedly higher GPU concurrency than buffered approaches; treat throughput claims as vendor numbers.
Part of the open Nemotron family
Nemotron is NVIDIA's family of open models — released with open weights, training data and recipes, and available as NVIDIA NIM microservices for self-hosting.
Server-side via the fal API
This page calls the current NVIDIA Nemotron ASR fal endpoint server-side so your FAL_KEY stays private and browser users only receive the transcript.
Transcribe with Nemotron in four steps
Everything happens in the demo above — just sign in with a free account to start. Each run costs 1 credit.
Add your audio
Upload a short clip (mp3, wav, ogg, m4a or aac, up to 10 MB) or paste a public audio URL into the Nemotron demo at the top of this page.
Pick language and acceleration
Leave language on auto-detect or choose a supported locale, then pick none, regular, high or full acceleration depending on the accuracy-latency trade-off you want.
Run the transcription
Press Transcribe. The request is sent to our server, which calls the fal Nemotron 3.5 ASR endpoint with your key kept private, then returns the text.
Copy or refine the result
Read the punctuated transcript, copy it, and adjust acceleration or input if you want to compare accuracy versus speed on the same clip.
Choosing an acceleration level
Acceleration sets the streaming chunk size. Smaller chunks mean lower latency; larger chunks mean more context and better accuracy.
Best accuracy. Use for final transcripts, recordings and anything you'll publish.
Balanced latency and accuracy — the default choice for most demo clips.
Faster responses for interactive use, with a clearer accuracy trade-off.
Lowest latency for live captioning and voice agents; expect the largest accuracy trade-off.
Chunk sizes are approximate and follow NVIDIA's documented configurations. fal currently shows pricing at roughly $0.008 per minute; check fal's API docs for the current schema and rates.
What you can build with Nemotron 3.5 ASR
Low-latency, multilingual speech recognition fits a wide range of audio products.
Live captions & subtitles
Stream speech into readable, punctuated captions for meetings, webinars and broadcasts where latency matters.
Voice agents & assistants
Feed low-latency transcripts into conversational AI so assistants can respond while the user is still speaking.
Multilingual transcription
Transcribe audio across roughly 40 language-locales from a single model without juggling per-language checkpoints.
Call & meeting analytics
Turn calls and meetings into searchable, punctuated text for QA, compliance and summarization pipelines.
Media & podcast workflows
Draft transcripts for editing, show notes and clip discovery, then refine in a full transcription workspace.
Accessibility tooling
Provide real-time speech-to-text for accessibility, note-taking and dictation experiences.
Current fal endpoint. This demo uses the nvidia/nemotron-asr-multilingual/asr endpoint for hosted Nemotron ASR transcription. Provider errors, timeouts or account configuration issues are refunded in the demo route. The Nemotron 3.5 ASR weights are open on Hugging Face for self-hosting, and you can always transcribe in the Whisper AI workspace .
Nemotron 3.5 ASR frequently asked questions
Common questions about Nemotron, Nemotron AI and Nemotron 3.5 ASR.
What is Nemotron 3.5 ASR?
Nemotron 3.5 ASR is NVIDIA's open, 0.6-billion-parameter multilingual streaming speech-recognition (ASR) model. It transcribes audio to text in real time using a Cache-Aware FastConformer-RNNT architecture, covers roughly 40 language-locales in a single checkpoint, and outputs native punctuation and capitalization.
What is Nemotron?
Nemotron is NVIDIA's family of open AI models, released with open weights, training data and recipes. The family spans language and speech models and ships as NVIDIA NIM microservices. Nemotron 3.5 ASR is the speech-recognition member used on this page.
Is Nemotron AI free to try here?
You'll need a free Whisper AI account to run the demo, and each transcription costs 1 credit. New accounts come with free starter credits so you can try Nemotron right away, and you can top up anytime on the pricing page.
How many languages does Nemotron 3.5 ASR support?
A single 0.6B checkpoint supports roughly 40 language-locales — about 36 languages — using language-ID prompt conditioning, so you don't need a separate model per language.
What does the acceleration setting do?
Acceleration controls the streaming chunk size and therefore the latency/accuracy trade-off. none uses ~1.12s chunks for the best accuracy; regular, high and full use progressively smaller chunks for lower latency.
What audio formats and sizes work?
fal documents mp3, ogg, wav, m4a and aac for the file input, and a data URI is accepted. This demo keeps browser uploads to about 10 MB; for longer files, host them and paste a public URL instead.
Is my audio private?
Your FAL_KEY never reaches the browser — requests go through our server. The audio itself is sent to NVIDIA's hosted endpoint on fal for processing, so avoid uploading confidential recordings to the public demo.
Why does the demo sometimes say the endpoint is unavailable?
The demo depends on fal's hosted NVIDIA Nemotron ASR endpoint, your account credits and our server-side FAL_KEY configuration. If the provider is unavailable or a request times out, the demo links you to the Whisper AI workspace so you can still get your audio transcribed.
Try Nemotron AI speech-to-text
Upload a clip, pick an acceleration level and read a punctuated transcript in seconds — then explore the open Nemotron weights to build your own pipeline.
Open weights · 0.6B parameters · ~40 language-locales