Nemotron 3.5 ASR — Streaming Speech to Text from NVIDIA
Put your own audio through Nemotron AI right here
Nemotron is the name NVIDIA gives its open model family; the 3.5 ASR member is the 0.6B multilingual streaming recogniser. Drop a clip in below, or paste an audio URL, and the punctuated text comes back in seconds.
Leave this on auto unless you already know what is being spoken.
More acceleration means smaller streaming chunks and less waiting; none stays at 1.12s chunks and is the most accurate, while full drops to 0.08s chunks and is the fastest.
to try it. New accounts include starter credits; after those, each run costs 1 credit.
Worth knowing: your audio goes to NVIDIA's hosted Nemotron 3.5 ASR endpoint on fal, relayed by our server. Should that provider be down, the same audio can be transcribed in the GPT Transcribe workspace.
Our own write-up. Nemotron and Nemotron 3.5 ASR belong to NVIDIA; the demo runs through fal's hosted nvidia/nemotron-asr-multilingual/asr endpoint.
What Nemotron 3.5 ASR is
600 million parameters, multilingual, and built to transcribe while the audio is still arriving. It is the speech entry in NVIDIA's Nemotron family, which ships with open weights, training data and recipes, and is packaged as NVIDIA NIM microservices.
The architecture is Cache-Aware FastConformer-RNNT with language-ID prompt conditioning — which is how a single checkpoint stretches to roughly 40 language-locales while still producing punctuation and capitalisation of its own.
The panel above is the fastest way to form an opinion. Requests go through our server to the hosted fal endpoint so no API key reaches your browser — and if you would rather run it yourself, the weights are on Hugging Face.
What it does well
Streaming output, one checkpoint for many languages, and latency you can change per request — a practical combination rather than a benchmark-winning one.
Transcribes as it listens
This is a streaming recogniser first. NVIDIA reports sub-100ms time-to-final on short chunks; that is their own benchmark, but the design intent — live captioning, voice agents — is clear enough from the architecture.
Cache-Aware FastConformer-RNNT
A Cache-Aware FastConformer encoder feeding an RNNT decoder, which means streaming context is carried forward rather than the buffer being re-processed at every step.
One checkpoint, many languages
Roughly 40 language-locales — about 36 languages — out of one 0.6B checkpoint, selected by language-ID prompt conditioning. No swapping models per language.
Punctuation comes built in
Output arrives already punctuated and capitalised, which removes the restoration pass that older pipelines needed before the text was readable.
Latency you set per request
Four chunk sizes — about 1.12s, 0.56s, 0.32s and 0.08s — selectable through the acceleration setting on each request. Nothing has to be redeployed to change the trade-off.
Small enough to serve widely
600M parameters keeps the per-stream cost low, and NVIDIA cites substantially higher GPU concurrency than buffered approaches achieve. Those throughput figures are the vendor's own.
Open, like the rest of Nemotron
The family ships open weights, training data and recipes, and is distributed as NVIDIA NIM microservices for anyone self-hosting.
Called from our server, not yours
The fal endpoint is invoked server-side, so the FAL_KEY never leaves our infrastructure and the browser receives nothing but the transcript.
Four steps to a transcript
All of it happens in the demo above. A free account is required; each run costs 1 credit.
Bring in the audio
A short clip — mp3, wav, ogg, m4a or aac, up to 10 MB — or a public audio URL, into the demo at the top of the page.
Choose language and speed
Auto-detect handles most clips; name the locale when you know it. Then decide where you sit between accuracy and latency: none, regular, high or full.
Send it
Hit Transcribe. Our server takes it from there, calls the fal Nemotron 3.5 ASR endpoint with the key held server-side, and hands the text back.
Read it, then try again
Copy the transcript if it is what you needed. If not, change the acceleration and run the same clip again — that comparison is the quickest way to find your setting.
How much latency to trade away
Acceleration is chunk size by another name. Smaller chunks return sooner; larger chunks give the model more context to be right with.
The accurate setting. Use it for finished transcripts and anything being published.
The middle ground, and the sensible default for trying a clip.
Quick enough to feel interactive, with the accuracy cost becoming noticeable.
As fast as it goes — for live captions and voice agents, and with the accuracy cost to match.
These chunk sizes are approximate and mirror NVIDIA's documented configurations. fal lists roughly $0.008 per minute at the time of writing — their API docs are the authority on both schema and price.
Where it fits in a product
Recognition that is both fast and multilingual opens up more product surfaces than either property alone.
Captions in real time
Meetings, webinars and broadcasts where captions have to keep up with the speaker rather than trail them.
Assistants that answer promptly
Transcripts arriving fast enough that a conversational system can begin forming a reply before the user has finished.
Transcribing across languages
Around 40 language-locales served by one model, which removes the routing layer that per-language checkpoints require.
Analysing calls and meetings
Punctuated, searchable text is what QA review, compliance retention and summarisation pipelines all need as their input.
Media and podcast production
A draft transcript is enough to start editing, write show notes, and find the clip worth cutting — refine it properly afterwards.
Accessibility features
Real-time text for assistive use, note-taking, and dictation built into a product.
The endpoint behind this demo. Transcription here runs through the nvidia/nemotron-asr-multilingual/asr endpoint for hosted Nemotron ASR. If the provider errors, times out, or the account is misconfigured, the demo route refunds the credit. The weights themselves are open on Hugging Face if you would rather host it, and there is always the GPT Transcribe workspace .
Nemotron 3.5 ASR, answered
The model, the wider Nemotron family, and how this demo works.
What Nemotron 3.5 ASR is
NVIDIA's open 0.6-billion-parameter multilingual ASR model. It works in real time on a Cache-Aware FastConformer-RNNT architecture, reaches roughly 40 language-locales from one checkpoint, and punctuates and capitalises the output itself.
What is Nemotron?
The umbrella name for NVIDIA's open models — weights, training data and recipes all published, delivered as NVIDIA NIM microservices. It covers language and speech; 3.5 ASR is the speech-recognition member this page demonstrates.
Is Nemotron AI free to try here?
A free account is needed and each run costs 1 credit — but new accounts arrive with starter credits, so trying it costs nothing. Top-ups are on the pricing page when you want more.
How many languages does Nemotron 3.5 ASR support?
Roughly 40 language-locales, which works out at about 36 languages, all from one 0.6B checkpoint via language-ID prompt conditioning. No per-language model to deploy.
What does the acceleration setting do?
It sets the streaming chunk size, which is the same thing as setting the latency/accuracy balance. none holds ~1.12s chunks for the most accurate result; regular, high and full shrink them in turn.
What audio formats and sizes work?
fal lists mp3, ogg, wav, m4a and aac, and will take a data URI. Browser uploads here are capped near 10 MB — anything longer should be hosted somewhere and passed as a public URL.
Is my audio private?
The key stays on our server and never reaches the browser. The audio, though, does travel to NVIDIA's hosted endpoint on fal to be processed — so keep confidential recordings out of this demo.
Why does the demo sometimes say the endpoint is unavailable?
Three things have to line up: fal's hosted endpoint, credits on your account, and our server-side FAL_KEY. When the provider is down or a request times out, the demo points you at the GPT Transcribe workspace so the audio still gets transcribed.
Give it something to transcribe
A clip, an acceleration level, and a punctuated transcript moments later. If it holds up, the open Nemotron weights are there to build a pipeline on.
Open weights · 0.6B parameters · ~40 language-locales