NVIDIA Nemotron 3.5 ASR: languages, speed modes, and honest limits

What Nemotron 3.5 ASR actually is

Nemotron 3.5 ASR is NVIDIA's automatic speech recognition model for streaming transcription. The published checkpoint has 0.6B parameters and is distributed through Hugging Face as nvidia/nemotron-3.5-asr-streaming-0.6b as of 2026. That size is deliberate: it sits far below multi-billion-parameter speech models because it is built to run fast in chunked, low-latency pipelines, not to win open-ended benchmark contests.

The architecture pairs a FastConformer encoder with an RNN-Transducer (RNNT) decoder. The encoder turns incoming audio into acoustic representations; the transducer emits text tokens as frames arrive. That token-by-token emission is what makes streaming output possible without waiting for a complete recording.

Output is plain text with punctuation and capitalization produced natively by the model. There is no separate cleanup pass bolting periods and capitals on afterwards, which tends to keep the punctuation consistent with how the audio was actually spoken.

How cache-aware streaming works

Streaming ASR models process audio in chunks. A naive implementation re-encodes overlapping context with every new chunk, paying for the same audio repeatedly and adding delay. Nemotron 3.5 ASR is cache-aware: the encoder keeps its intermediate states from earlier chunks and reuses them, so each new chunk only spends compute on the audio it actually adds.

Two practical consequences follow. First, latency stays low and roughly constant as a recording grows, instead of creeping up with length. Second, chunk size becomes a genuine tuning knob: you decide how much context the encoder sees per step, and that decision trades accuracy against responsiveness.

Chunk sizes and the four speed modes

The model exposes four chunk-size settings. Larger chunks give the encoder more context per step, which helps accuracy; smaller chunks shrink the wait between speech and text. The free transcriber on this site maps them to four named modes:

If you are unsure, start with Balanced. Move to Best accuracy when a transcript misses names or domain terms, and reach for the low-latency settings only when turnaround time genuinely matters more than the occasional dropped word.

All 40 supported languages

Nemotron 3.5 ASR accepts exactly 40 locales, listed below grouped roughly by region. Chinese means Mandarin (zh-CN). This is the same whitelist the transcriber on this site enforces, so the table and the tool can never drift apart.

Auto-detection and mid-recording code-switching

You do not have to declare a language up front. Leave the setting on auto and the model identifies the language from the audio itself. Detection also holds when a speaker switches languages partway through: start a voice note in English, drift into Spanish, and the transcript follows without a restart.

Explicitly picking a locale still helps on short clips and heavy accents, simply because auto-detection has less audio to work with. If you know the recording is Mandarin, set zh-CN and remove the guesswork.

What Nemotron 3.5 ASR does not do

An honest limits list, because most ASR marketing skips it:

• No speaker diarization. Everyone's speech lands in one continuous transcript. Pair it with a dedicated diarization model if you need who-said-what labels.
• No verified word or segment timestamps as of 2026, which also means no native SRT or VTT subtitle export from its plain-text output.
• Audio in, text out. Video files need their audio track extracted before transcription.
• In browser use, the transcript appears after the recording or upload is processed. The streaming architecture keeps that wait short, but it is not a live caption overlay.

Try Nemotron 3.5 ASR free in your browser

You can test the model on this site without installing NeMo, provisioning a GPU, or creating an account. nemotronasr.com is independent and not affiliated with NVIDIA; it provides browser access to the model through a hosted endpoint.

The free tier covers 3 transcriptions a day with no sign-up. New Pro purchases are paused; existing license keys remain supported.

1. Open the transcriber on the homepage.
2. Upload an MP3, WAV, M4A, AAC, OGG, or OPUS file (up to 5 minutes and 20 MB), or record straight from your microphone.
3. Leave language on auto-detect, or pick one of the 40 locales.
4. Choose a speed mode. Balanced is the safe default.
5. Copy the punctuated transcript, or download it as a .txt file.