Whisper, Parakeet and Nemotron Transcribe on a DGX Spark: local speech to text in November 2026

On a DGX Spark, NVIDIA's Parakeet TDT 0.6B v3 transcribes at 283 times real time and Whisper large-v3 at 19.5 to 24.8 times real time, according to one owner's January 2026 benchmark across six engines on a GB10. On the Hugging Face Open ASR Leaderboard (updated 6 November 2026), Parakeet v3 averages 7.04% English WER against 8.24% for Whisper large-v3, and the best open model, NVIDIA's Nemotron 3.5 Transcribe 30B-A4B, 6.13%. At those speeds an hour of audio takes Parakeet about 13 seconds, and the common models need 2 to 28 GB of memory, so they fit next to an LLM on the same 128 GB box.

Source: Open ASR Leaderboard, "Latest results", last updated 6 November 2026, read 7 November 2026. Sizes, languages and licenses from each model card. The English average covers nine datasets, two of them private held-out sets from Appen and DataoceanAI. RTFx (seconds of audio per second of compute) has been measured on an H200 since 24 June 2026, so it ranks models against each other and does not predict GB10 speed. Long-form WER is the leaderboard's Earnings21/Earnings22/CORAAL average.

Which open speech model is most accurate?

Nemotron 3.5 Transcribe and Qwen3-ASR-1.7B, separated by 0.02 points of WER. Both are newer than Whisper, and both beat it clearly on English.

The leaderboard ranks 63 models, closed APIs included. The best of those, Zoom's scribe v2.1 pro, averages 4.87%; the best open model sits ninth. NVIDIA's own Nemotron 3.5 Transcribe model card (released 28 October 2026) claims 4.71% on "the English Open ASR nine-dataset mean", 1.4 points better than the leaderboard's own run; I use the leaderboard figure because the same evaluation script scored every row. Two other points stand out. Whisper large-v3-turbo, with its decoder cut from 32 layers to 4 according to its model card, costs half a point of WER against large-v3. And a 0.6B Parakeet beats both Whisper models on English while running about eight to thirteen times faster on the same H200.

For other languages the picture changes. On the leaderboard's multilingual tab (German, French, Italian, Spanish, Portuguese, Hindi, Dutch), Voxtral Small 24B averages 6.08% and Whisper large-v3 8.19%, while Parakeet v3 and Canary 1B v2 go unranked because they cover six of the seven languages. Parakeet v3's 25 languages include Czech, Polish and others the multilingual tab does not test, per the Parakeet v3 model card.

How fast is transcription on a DGX Spark?

Parakeet runs at a few hundred times real time and Whisper large-v3 at about 20 to 25 times. The one systematic GB10 benchmark I found is coder543's post in the NVIDIA forum thread Running Parakeet speech to text on Spark (7 January 2026), which ran six engines on the same English sample.

Source: coder543 in NVIDIA forum thread 356353, 7 January 2026, GB10 with 119.7 GiB visible, one English file (jfk_rice_16k.wav), mean of 3 runs for whisper.cpp and 5 for the rest, batch size not given. WER on one short file says little about accuracy; use the leaderboard for that. Engine versions were not stated.

Worked through for an hour of audio, by my arithmetic: Parakeet v3 needs 3,600 / 282.9 = 12.7 seconds, WhisperX large-v3 145 seconds and whisper.cpp large-v3 185 seconds. At $0.79 per hour of Spark time, 1,000 hours of recordings cost about $2.80 with Parakeet (3.5 hours of machine time) and about $40 with whisper.cpp (51 hours). These are single-file rates; NeMo and vLLM batch requests, which is how the H200 leaderboard reaches 6,322x for Parakeet, so a batched job on the Spark should do better than this table. For a queue of files, the batch processing guide covers job layout and checkpointing.

A slower end of the range: in an October 2026 forum post, Microsoft's 8.7B VibeVoice-ASR in BF16 Transformers took 85 minutes 36 seconds for a 59:24 press conference on a Spark, slower than real time (RTF 1.44), with a 28.1 GB peak. A 10-minute excerpt took 367 s on the Spark against 80 s on an RTX 4090; the author suspects bandwidth (273 vs 1,008 GB/s) but did not profile it. That model generates text token by token, so unlike Parakeet it pays the LLM decode cost.

Which serving option works on ARM64?

NeMo in a container, whisper.cpp built from source, vLLM, or NVIDIA's Riva ASR NIM. Pip-installed wheels are where most owners lose time.

• NeMo (Parakeet, Canary): use Docker. In the Parakeet thread, a pip-installed NeMo could not see CUDA, and mattdee traced a Lhotse error in the NGC PyTorch 25.12 container to PyTorch 2.10; the workaround was NGC PyTorch 25.10 with PyTorch 2.9. coder543 recommends Docker "for just about everything" and ships a CUDA 13 Dockerfile.
• faster-whisper and WhisperX: the CTranslate2 install page says "The prebuilt aarch64 wheels on PyPI are CPU-only", so GPU inference needs a source build. One owner in the whisper.cpp thread ran WhisperX through a Blackwell-specific build behind FastAPI instead.
• whisper.cpp: no ARM64 CUDA binaries are released. swann.schilling's May 2026 recipe builds v1.8.4 in docker.io/nvidia/cuda:13.0.3-devel-ubuntu24.04 with CUDA architectures "120;121"; ggml-large-v3-turbo then uses 1,624 MB of GPU memory.
• vLLM: a forum user reports Whisper large-v3 working through /v1/audio/transcriptions. The Qwen3-ASR card lists day-0 vLLM support (streaming only via vLLM, timestamps through a separate 0.6B forced aligner), and the Granite Speech 5.0 card lists vLLM, llama.cpp and Ollama. Nemotron 3.5 Transcribe names DGX Spark (GB10) in its hardware table with BF16, FP8 and NVFP4 on vLLM 0.32.0 plus the NeMo Speech plugin. Engine setup is in the vLLM install guide.
• Riva ASR NIM: NVIDIA's ASR NIM support matrix (updated 6 November 2026) lists Parakeet 1.1B CTC English, Parakeet 1.1B RNNT Multilingual and Nemotron ASR Streaming on DGX Spark at FP16, plus a Nemotron Transcribe NIM (nemotron-transcribe:1.0.0) with an ARM64 runtime. In January the forum reported the Parakeet NIM as x86-only; the matrix now says otherwise. Licensing is covered in the NIM on DGX Spark guide.

The shortest path I can point to for Parakeet is the model card's own NeMo call inside a CUDA 13 PyTorch container:

import nemo.collections.asr as nemo_asr

asr = nemo_asr.models.ASRModel.from_pretrained("nvidia/parakeet-tdt-0.6b-v3")
out = asr.transcribe(["meeting.wav"], timestamps=True)
print(out[0].text)

The card says full attention handles up to 24 minutes per file on an 80 GB A100 and local attention up to 3 hours; split long recordings or switch the attention mode.

How do you add speaker labels?

Either run a separate diarization model and merge timestamps, or use a model that emits speakers itself. The options with published numbers:

• pyannote community-1: CC-BY-4.0, gated on Hugging Face, for pyannote.audio 4.x, CPU by default and movable to CUDA. Its card reports 17.0% DER on AMI headset audio and 19.9% on the single distant mic, with no collar.
• Sortformer inside the Riva NIM: the Parakeet 1.1B CTC English NIM has Sortformer plus Silero VAD profiles at 7.23 to 47.22 GB.
• Nemotron 3.5 Transcribe: tags speakers as <spk:N> and returns word timestamps from CTC heads in vLLM; the card claims 15.80% mean cpWER on multi-speaker sets (self-reported).
• VibeVoice-ASR: one pass, speakers and timestamps, 11.3% DER on one 39-minute AMI meeting with headsets, but it split a press conference into 25 speakers and can loop on noisy audio until the token limit, per the forum post.

Can a Spark run speech to text and an LLM together?

Yes, memory is not the constraint. The typical pairing is a transcriber plus an LLM that summarizes or extracts from the transcript.

My Nemotron figures are 34B parameters times 2, 1 and about 0.56 bytes, weights only. The NIM figures are NVIDIA's per-profile numbers, not labeled per platform.

A Parakeet NIM at 5.83 GB next to Qwen3.6-35B-A3B at about 24 GB of NVFP4 weights leaves most of the 128 GB for KV cache. The trap is start order: vLLM's --gpu-memory-utilization is a fraction of total memory, so an LLM started at 0.9 claims the space the speech model needs. Start the ASR service first and cap the LLM; the multi-model memory guide walks through the budget. Both services share one GPU, so expect each to slow down while the other is busy; I found no published measurement of how much for ASR.

Which speech model should you run on a Spark?

For English and European-language batch transcription, Parakeet TDT 0.6B v3 in NeMo: 283x real time on GB10 in the one benchmark, 7.04% WER, CC-BY-4.0, and small enough to forget about. For the best open accuracy with speakers and timestamps in one model, Nemotron 3.5 Transcribe, which lists the Spark as a target, though nobody has published a GB10 speed for it yet. For languages outside Parakeet's 25, Whisper large-v3 rather than turbo: the turbo saves time you do not need on this hardware and costs half a point of WER. I would not use VibeVoice-ASR for long files on a Spark, since it runs slower than the audio. Recordings of identifiable people are usually personal data under the GDPR; the EU compliance page sets out what changes when it stays on dedicated hardware. For the other direction, text to speech, the open models are compared in open source TTS models.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• Open ASR Leaderboard, Hugging Face (hf-audio), 2026-11-06
• open_asr_leaderboard repository, GitHub (huggingface)
• Running Parakeet speech to text on Spark, NVIDIA Developer Forums, 2026-01-07
• Running whisper.cpp STT server on DGX Spark (GB10, ARM64 + CUDA 13) via Docker, NVIDIA Developer Forums, 2026-05-30
• VibeVoice-ASR on DGX Spark: a 59-minute recording in one pass, with speakers and timestamps, NVIDIA Developer Forums, 2026-10-07
• NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model card, Hugging Face (nvidia), 2026-10-28
• parakeet-tdt-0.6b-v3 model card, Hugging Face (nvidia)
• NVIDIA ASR NIM Support Matrix (26.10.0), NVIDIA Docs, 2026-11-06
• CTranslate2 installation, OpenNMT
• Qwen3-ASR-1.7B model card, Hugging Face (Qwen)
• granite-speech-5.0-2b model card, Hugging Face (ibm-granite), 2026-10-20
• whisper-large-v3-turbo model card, Hugging Face (openai)
• whisper-large-v3 model card, Hugging Face (openai)
• Voxtral-Small-24B-2507 model card, Hugging Face (mistralai)
• canary-qwen-2.5b model card, Hugging Face (nvidia)
• pyannote/speaker-diarization-community-1, Hugging Face (pyannote)