Nemotron 3.5 Transcribe: What It Is, How Accurate It Is, and How to Run It

NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026. It is an open-weight speech-to-text model for recorded audio that transcribes 40 languages, labels who said what, and translates between English and 31 languages (NVIDIA's launch post, model card). The weights ship under the OpenMDW-1.1 license. Artificial Analysis has the model tied for first among open-weight models at 2.8% AA-WER.

Nemotron 3.5 Transcribe is a separate model from Nemotron 3.5 ASR streaming, the 0.6B FastConformer checkpoint NVIDIA released in June 2026 for live audio. Transcribe is a 34B-parameter, LLM-based model, named 30B-A4B, and it takes a whole recording per request (Transcribe card, streaming card). Live voice agents still need a streaming model.

This guide covers what the model does, how four benchmarks score it, which languages it supports, how to run it, and how it compares with Whisper, Voxtral, and Parakeet. The numbers come from NVIDIA's model card and launch post, Artificial Analysis, the Hugging Face Open ASR Leaderboard, and our own Converse-STT benchmark, where we ran the model on real phone calls.

What Nemotron 3.5 Transcribe is (and what it is not)

Nemotron 3.5 Transcribe pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model, and you steer it with a text prompt. Sortformer speaker cues feed the language model alongside the acoustic features (NVIDIA's post). NVIDIA's post describes it as about 30B parameters with about 4B active. The model card counts 34B total and about 4.3B active including the audio encoder, and this guide uses the card's numbers.

The model card says it directly: "This is an offline model." Internal chunking of long audio does not turn it into a real-time service (model card).

Both models carry a "6x" claim, and the two are unrelated. For Transcribe, 6x is Artificial Analysis's active-parameter ratio against Voxtral Small, 4B versus 24B (Artificial Analysis). For the streaming model, 6x is a concurrency ratio at 1,120 ms on a single H100 (streaming card).

For live calls, pick the streaming model or another streaming STT. For a wider field of streaming options, see our guide to the best voice AI APIs for real-time audio processing. Our voice AI latency guide explains how STT timing fits into a full turn. Transcribe suits post-call transcription, QA, analytics, meetings, medical dictation, and batch jobs.

Launch facts at a glance

• Release: October 28, 2026, alongside Granary v2, labels and manifests for 463,062 hours of speech in 48 languages (NVIDIA's post).
• Language backbone: Nemotron 3.5 Lightning (model card).
• Weights: BF16 is public. The card names FP8 and NVFP4 variants, but their repos were not public when checked on Oct 28 (model card).
• Hosted endpoints: Baseten, DeepInfra, and Simplismart (NVIDIA's post).
• Hardware: The card lists Jetson Thor, DGX Spark, and RTX 5090 alongside H100 and Blackwell data-center GPUs. NVIDIA's post says quantized builds for "RTX Spark, DGX Spark, or Jetson AGX Thor" are coming soon, so treat the small-device entries as listed support until then.
• Throughput: 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell GPU, using the forthcoming NVFP4 build and 30-second clips at high concurrency (NVIDIA's post). The card ties its throughput figures to the Nemotron Transcribe NIM and says the open-source build is about two times slower.

Nemotron 3.5 Transcribe accuracy: four benchmarks, four answers

Four sources score this model, and each answers a different question. The table puts them side by side with the source and date on every row.

The two Artificial Analysis numbers differ. The post reports 2.8%. The leaderboard row measures the DeepInfra deployment and shows 3.4%. Artificial Analysis does not say why. The gap may reflect different endpoints, which is an inference. Hosting changes scores: the same Whisper Large v3 runs from 4.1% on fal.ai to 10.1% on Replicate in that table.

On the same leaderboard, Whisper Large v3 also scores 4.5% on Together AI, Parakeet TDT 0.6B v3 scores 4.5%, and Deepgram Nova-3 scores 5.2%. AssemblyAI Universal-3.5 Pro scores 3.0%, and ElevenLabs Scribe v2 scores 2.2%.

The Open ASR and Artificial Analysis numbers are both correct. They use different test sets. The Open ASR average covers 11 datasets, including a private conversational set where Nemotron scores 12.73%. AA-WER v2 gives voice-agent conversations 50% of the weight. The model card reports a nine-set English mean of 4.71, so name the set whenever you quote a figure.

The leaderboard ranks Nemotron 9th of 61 overall. NVIDIA calls it first among open-weight models, and its figure caption says API models are excluded. The eight rows above Nemotron list no RTFx, which suggests API systems, though that is an inference. The next rows with a published RTFx are Qwen3-ASR-1.7B at 5.40, granite-speech-5.0-2b at 5.68, cohere-transcribe at 5.77, Hojo-ASR-V1 at 5.85, thewhisper-large-v3-turbo at 5.91, higgs-audio-v3-stt and ARK-ASR-3B at 5.92, and ARK-ASR-0.6B at 5.93. Canary-Qwen 2.5B follows at 5.94, ranked 21st (Open ASR Leaderboard). Keep the open-weight qualifier whenever you quote the ranking.

NVIDIA's launch post adds three more results (NVIDIA's post):

• VoiceArena Diarization-Bench: 32.4% cpWER, first of 14 systems.
• David AI DAI-ASR-I18N: first among open-weight models.
• Eka Care medical speech: 30.8% WER out of the box, 23.8% with a better prompt, and 11.8% after fine-tuning.

What our benchmark shows on real phone calls

Clean test clips flatter speech models, so at Cekura we also score them on real calls. On the Converse-STT leaderboard, retrieved Oct 28, 2026 (the page's own stamp reads Sep 16, 2026), Nemotron 3.5 Transcribe posts 2.28% overall, the lowest of 21 models. It scores 1.89% on Pipecat clips and 3.03% on Ocular real calls, where Reson8 is lowest at 2.93%. Eight recordings make a small sample, so read the Ocular ranking as a close race. Real-call WER runs above clip WER, and likely causes are phone-band audio, crosstalk, and accents. Our guide to accent and dialect testing shows how to probe the accent part. Full results are in our full benchmark write-up. The compare page lists every head-to-head.

Prompting changes the transcript

The model takes a text prompt alongside the audio, and the prompt changes the output. The model card lists these patterns (model card):

• Verbatim: Audio transcription. Language: as spoken. Wording: verbatim, including fillers and repetitions. Formatting: punctuation and capitalization. Output: transcript only. Use it for compliance review and QA, where the fillers and repetitions matter.
• Clean: Audio transcription. Language: as spoken. Wording: omit fillers and accidental repetitions; keep the remaining spoken wording unchanged. Formatting: punctuation and capitalization. Output: transcript only. Use it for readable summaries and agent-assist text.
• Context terms: Transcribe with punctuation and capitalization. Context terms: <names and terms>. Use it for product names, drug names, and street addresses.
• Speaker tags: a prompt that asks for <spk:0> through <spk:3> with word timestamps enabled. Use it for two-party calls and conference lines.
• Near-field (experimental): a prompt that transcribes only the speaker nearest the microphone.
• Translation: Please translate the given speech to <target language>. Use it to read calls in a language you do not speak.

Prompt choice matters. In the Eka Care case study, a better prompt moved WER from 30.8% to 23.8% before any fine-tuning (NVIDIA's post). Our benchmark used the verbatim prompt unchanged, so its 2.28% reflects no prompt tuning.

Audio Q&A also exists. NVIDIA trained and evaluated it in English, and the card warns it may not work reliably in other languages. The model card lists known limits: the model can hallucinate speech, pick the wrong language, and mis-attribute speakers.

Which languages Nemotron 3.5 Transcribe supports

Nemotron 3.5 Transcribe covers 40 languages with ASR support. The model card sorts them into three quality levels. Leading languages have the best out-of-the-box quality, Production-ready languages reach production-level quality, and Supported languages may need fine-tuning (model card).

The card's Supported tier also lists Norwegian, but its matrix shows no ASR support for that row. This guide leaves Norwegian out and counts 40. Translation runs in both directions between English and 31 languages. Speaker labels work in 13 languages: English, French, German, Hindi, Italian, Korean, Portuguese, Russian, Spanish, Japanese, Thai, Arabic, and Chinese Mandarin. The streaming model uses its own three tiers of 40 locales: 19 transcription-ready, 13 broad-coverage, and 8 adaptation-ready (streaming card).

Speaker attribution, translation, and long audio

Speakers. Nemotron 3.5 Transcribe adds speaker diarization to the transcript, with labels in 13 languages. NVIDIA's post, TechCrunch, and The Next Web say up to eight speakers. The model card states no formal limit, shows tags <spk:0> to <spk:3> in its example prompt, and says training mixtures included conversations of up to eight speakers (model card). Treat eight as the ceiling seen in training and test your own calls. With word timestamps on, the response carries per-word start, end, and speaker fields under ctc_timestamps. Labels separate voices in a recording and establish no real-world identity.

Translation. Translation is unavailable on the SGLang-Omni path, which also omits diarization and word timestamps (model card).

Long audio. One recording can run up to 1 hour. vLLM caps audio decoding at 600 seconds by default, so set VLLM_MAX_AUDIO_DECODE_DURATION_S=3600 before you send long files. Split anything longer than an hour on your side. On the long-form Open ASR set, Nemotron's mean WER is 10.98% against 13.37% for Parakeet-TDT-0.6B-v3 (model card).

How to run Nemotron 3.5 Transcribe

Run it on Linux or WSL2 with one H100 80 GB, the card's reference setup, plus vLLM 0.32.0 and the NeMo Speech plugin. Download the weights with hf download nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B --local-dir ./NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. NVIDIA's NeMo Speech repository on GitHub supplies the plugin and fine-tuning tooling. A demo Space on Hugging Face lets you try it in a browser. The code below is copied from the model card as of Oct 28, 2026, not from a run.

export VLLM_PLUGINS=nemo_speechlm
export VLLM_MAX_AUDIO_DECODE_DURATION_S=3600
export VLLM_MAX_AUDIO_CLIP_FILESIZE_MB=180
export VLLM_USE_V2_MODEL_RUNNER=0   # word timestamps need the V1 runner
vllm serve ./NVIDIA-Nemotron-3.5-Transcribe-30B-A4B \
  --served-model-name nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B \
  --trust-remote-code --dtype bfloat16 --gpu-memory-utilization 0.94 \
  --max-model-len 65536 --max-num-batched-tokens 4096 --max-num-seqs 192 \
  --enable-chunked-prefill --no-enable-prefix-caching \
  --use-replayssm --mamba-backend flashinfer --mamba-cache-mode none \
  --mm-processor-cache-gb 0 --no-async-scheduling \
  --limit-mm-per-prompt '{"audio":{"count":6,"length":120000000}}' \
  --speculative-config '{"method":"mtp","num_speculative_tokens":8}' \
  --reasoning-parser nemotron_v3 \
  --trust-request-mm-kwargs \
  --middleware nemo.collections.speechlm2.vllm.salm.ctc_serving.ctc_timestamp_middleware

import base64
from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")
audio = base64.b64encode(open("audio.wav", "rb").read()).decode()
response = client.chat.completions.create(
    model="nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B",
    messages=[{"role": "user", "content": [
        {"type": "text", "text": "Transcribe the following audio."},
        {"type": "audio_url", "audio_url": {"url": f"data:audio/wav;base64,{audio}"}},
    ]}],
    temperature=0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}, "skip_special_tokens": False},
)
print(response.choices[0].message.content)

The card gives three caveats:

• Speed: this open build is intended for model development and is about two times slower than the NIM.
• Warm-up: the first requests are slow while kernels compile and warm up.
• Long files: a dense hour of audio can need about 20,000 output tokens. The request above sets no max_tokens, so raise it before you send hour-long files.

Other routes exist (model card). Hugging Face Transformers runs it in-process after a source install. SGLang-Omni 0.2.0 serves whole-file transcription. GGML is listed for the RTX 5090 and RTX PRO 6000. The card also lists A100, H200, L40S, B200, GB200, GB300, DGX Spark, and Jetson Thor.

What it costs against hosted APIs

Self-hosting replaces the per-minute rate with your GPU time, and NVIDIA publishes no list price for it. The table compares DeepInfra's hosted rate with other APIs, per hour of audio. For a closer look at the Deepgram line, see our Deepgram pricing breakdown.

License. The model card states that the OpenMDW License Agreement, version 1.1 governs the model, and that the model "is ready for commercial or non-commercial use" (model card). The Next Web says the license allows commercial use. We did not review the license text, so read it at openmdw.ai for attribution and patent terms before you ship.

Nemotron 3.5 Transcribe vs Whisper, Voxtral, and Parakeet

Blank cells mark values that could not be confirmed from a published source. WER figures come from different benchmarks, so compare within a column.

Sources: Artificial Analysis, Open ASR Leaderboard, Voxtral Small card, Mistral docs, OpenAI pricing, and the Whisper README.

Against Whisper, Nemotron wins on published accuracy and native speaker labels. Its AA-WER of 3.4% on DeepInfra sits below Whisper Large v3's 4.1% on fal.ai and 4.5% on Together AI. On the Open ASR Leaderboard, Nemotron averages 5.33 and whisper-large-v3 averages 7.16. Whisper reads recordings in a sliding 30-second window, the 0.6B streaming model works in 80 ms to 1,120 ms chunks, and Transcribe is the one that competes with Whisper on recorded audio. Whisper keeps two advantages no benchmark captures. whisper.cpp runs Whisper on CPUs, and whisper_streaming adds live transcription. OpenAI has deprecated the hosted whisper-1 API, with removal set for Feb 26, 2027. Check your pipeline's tools before you migrate.

Against Voxtral Small, Nemotron ties at 2.8% AA-WER in Artificial Analysis's post with about 4B active parameters against Voxtral's 24B. The leaderboard row puts Voxtral Small ahead, 2.8% against 3.4%. Against Parakeet, the Open ASR long-form set shows Nemotron at 10.98% against 13.37% for Parakeet-TDT-0.6B-v3. Parakeet is the smaller model, and the Nemotron 3.5 ASR streaming checkpoint is the live-audio option from the same family.

Testing the model before you switch your agent to it

Use Transcribe for recorded audio, keep a streaming model in the live path, and check either choice on your own calls before you switch. Swapping the STT model in a production voice agent changes more than word error rate. Transcript timing shifts turn-taking, and a new model makes different mistakes on names, numbers, and accents. Run the same scenario set against the old and new stack, and compare WER and other voice AI evaluation metrics on your own calls.

Cekura covers that comparison in three buckets:

• Pre-production simulations: run thousands of caller scenarios against the old and new STT setups and compare the results.
• Infrastructure testing: replay interruptions, background noise, and poor audio, where transcription errors start.
• Production call QA: score live calls for drop-off points, sentiment, workflow adherence, and compliance failures.

Native integrations work out of the box for Retell, VAPI, ElevenLabs, LiveKit, Pipecat, Bland, and more. You don't rebuild anything. You add a testing and monitoring layer on top of what you already have.

Cekura supports SOC 2, HIPAA, and GDPR compliance, covering transcript redaction, role-based access, and audit trails.

Book a demo to run that comparison on your own calls.

FAQs

What is the NVIDIA speech-to-text model?

NVIDIA has several speech-to-text models. Nemotron 3.5 Transcribe is the offline, LLM-based model for recorded audio, with translation and speaker labels. Nemotron 3.5 ASR streaming 0.6B handles live audio in 80 ms to 1,120 ms chunks. Parakeet and Canary are NVIDIA's older open ASR models.

Is Nemotron 3.5 Transcribe better than Whisper?

Yes, on published accuracy benchmarks. Artificial Analysis lists Nemotron 3.5 Transcribe at 3.4% AA-WER on DeepInfra and 2.8% in its own post, against 4.1% to 10.1% for Whisper Large v3, depending on the host. Nemotron also adds speaker labels and richer prompting. Whisper runs on CPUs through whisper.cpp, and Nemotron works on recorded audio only, so test both on your own calls.

What languages does Nemotron 3.5 Transcribe support?

Nemotron 3.5 Transcribe covers 40 languages with ASR support, translates between English and 31 languages, and labels speakers in 13. The model card sorts languages into Leading (13), Production-ready (15), and Supported tiers, so check your locale there. The streaming model covers 40 locales in three tiers: 19 transcription-ready, 13 broad-coverage, and 8 adaptation-ready.

Is Nemotron 3.5 Transcribe free?

The weights are free to download under the OpenMDW-1.1 license, and the model card says the model is ready for commercial or non-commercial use. Running the model costs compute. DeepInfra hosts it at $2 per 1,000 minutes, about $0.12 per hour, and self-hosting needs a GPU such as the card's 1x H100 80 GB reference setup. Read the license text before you ship.

Does Nemotron 3.5 Transcribe do real-time transcription?

No. Nemotron 3.5 Transcribe is an offline model that takes a recording of up to one hour per request. For live calls, use a streaming model such as Nemotron 3.5 ASR streaming 0.6B.

What is the difference between Nemotron 3.5 ASR and Nemotron 3.5 Transcribe?

The main difference between the two is mode and size. Nemotron 3.5 ASR streaming is a 0.6B FastConformer model that handles live audio in 80 ms to 1,120 ms chunks. Nemotron 3.5 Transcribe is a 34B LLM-based model that processes whole recordings and adds translation, speaker labels, and prompting.

How accurate is Nemotron 3.5 Transcribe on phone calls?

The Converse-STT benchmark measured 3.03% WER on 8 real phone recordings and 2.28% overall across clips and calls, the lowest overall of 21 models. Reson8 is lowest on real calls at 2.93%. The sample is small, so test your own calls.