Transcribe live multilingual speech with latency control

Nemotron 3.5 ASR is an open automatic speech recognition model from NVIDIA. It transcribes speech into text, adds punctuation and capitalization, and can label the detected language when you run it in automatic language mode.

The useful angle is streaming. Instead of waiting for a full recording to finish, the model is built to process audio in chunks. You can choose shorter chunks when you care about lower delay, or longer chunks when you want to give the model more context. The model card lists chunk sizes from 80ms to 1.12s, which makes the tradeoff easy to test instead of treating latency as a fixed property.

It also matters that this is one multilingual model rather than a separate setup per language. The model card lists 40 language-locales, with 32 producing transcription out of the box and another 8 intended for fine-tuning. That makes it a better fit for voice-agent logs, meeting capture, multilingual support calls, or any workflow where language detection is part of the job.

There is no public Space for a one-click browser test, so start from the model card and the NVIDIA NeMo streaming example. Install NeMo, load `nvidia/nemotron-3.5-asr-streaming-0.6b`, and run the cache-aware streaming inference script on a short WAV file.

For the first useful test, do not use a clean single sentence. Use a 30-60 second clip with a real speaker, pauses, names, and at least one sentence boundary. If you have multilingual audio, run one pass with a known `target_lang` and another with `target_lang=auto` so you can see whether the language tag is useful.

Then repeat the same clip with two chunk settings: one short setting for low delay and one longer setting for more context. Compare three things: whether the words are right, whether punctuation lands in sensible places, and whether the lower-latency setting is good enough for your use case.

This is not the easiest audio model to try casually. Local use expects a NeMo setup, Linux-oriented dependencies, and NVIDIA GPU tooling for realistic performance. Also, not every listed language is equally ready: the model card separates out 8 adaptation-ready locales that need fine-tuning for full transcription quality.

• Prototype live captions for meetings, calls, streams, or support workflows.
• Test multilingual transcription without wiring together separate language-specific models.
• Compare low-latency and higher-context settings on the same audio.
• Label mixed-language speech while keeping a readable transcript.
• Build a local transcription path before deciding whether a hosted ASR API is worth using.

Try the demo

View model page