Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time — we’ll notify you if affected. OEM costs and supply tightening — contact us to secure stock and pricing. For projects or bulk orders, choose “Request for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

AI & Workstations

Can a GB10 Transcribe 100 Hours of Meetings? Let's Put It to the Test.

Yes, comfortably, and probably in an afternoon. Published speed measurements on GB10 hardware put OpenAI's Whisper large-v3 at roughly 19 to 25 times faster than real time and NVIDIA's Parakeet models at roughly 280 to 460 times. On those figures, 100 hours of recordings would take about four to five hours with Whisper, or well under an hour with Parakeet, all on a desk-sized system with no audio leaving your office. We have not run this test ourselves: the speeds below come from a user benchmark on NVIDIA's developer forum, measured on one short sample, so treat our totals as estimates and check them on your own recordings. The ASUS Ascent GX10 starts at S$8,100 and the NVIDIA DGX Spark is S$8,800 at SourceIT (listed 28 Oct 2026).

How fast is speech-to-text on a GB10?

Model and tool Speed vs real time (published) Estimated time for 100 hours of audio
NVIDIA Parakeet-CTC 0.6B About 458x About 13 minutes
NVIDIA Parakeet-CTC 1.1B About 432x About 14 minutes
NVIDIA Parakeet-TDT 0.6B v3 About 283x About 21 minutes
Whisper large-v3, WhisperX or Transformers About 25x About 4 hours
Whisper large-v3, whisper.cpp on GPU About 19x About 5 hours 15 minutes

Speeds from a user benchmark on NVIDIA's developer forum, January 2026, on a single short audio sample. Estimates are simple arithmetic (100 hours divided by the speed factor) and exclude file handling, speaker labelling and summarising. NVIDIA has no official transcription playbook for DGX Spark.

Whisper or Parakeet: which should you use?

Speed is not the only consideration. Whisper is trained on many languages, including Mandarin, Malay and Tamil, and handles mixed-language speech reasonably well, which matters for Singapore meetings where people switch languages mid-sentence. NVIDIA's Parakeet models are much faster but are primarily built for English, with some versions adding other languages; check the model card for each language you need. A practical approach is Parakeet for English-only recordings and Whisper for multilingual ones.

Accuracy on your own audio matters more than any benchmark. Accents, crosstalk, poor microphones and room echo hurt every model. Our post Why Your Next Meeting Room Upgrade Might Be a Microphone, Not a Camera explains why the recording often decides the transcript quality.

Transcription speed is no longer the bottleneck. Recording quality and speaker labels are.

What else does a meeting-transcription job need?

How much storage does 100 hours need?

Less than you might expect. Uncompressed 16 kHz mono audio, the format speech models typically use, is about 115 MB per hour, so 100 hours is roughly 11.5 GB; compressed recordings are far smaller. The transcripts are text: 100 hours of speech is on the order of a million words, a few megabytes. Even the 1 TB GB10 configurations hold years of meeting audio. Storage choices are covered in 1TB vs 2TB vs 4TB NVIDIA GB10.

Why transcribe locally rather than in the cloud?

Board meetings, HR discussions, client calls and medical or legal conversations are exactly the recordings organisations are least comfortable sending to an external service. A local system keeps the audio, transcripts and summaries on your network. Cost is the other reason: cloud transcription is usually priced per minute of audio, while a GB10 that already exists for other AI work adds transcription at no extra charge. The trade-off is that you own the setup, the updates and the quality checks. The broader local-versus-cloud case is in Stop Paying for Cloud AI? Here's When Running AI Locally Makes Sense.

Run your own 100-hour test

  1. Collect a representative sample: different rooms, speakers and languages.
  2. Pick two models, for example Whisper large-v3 and a Parakeet model.
  3. Time one hour of audio end to end, including speaker labelling if you need it.
  4. Have someone check a few minutes of each transcript against the recording.
  5. Multiply up to your volume and add time for summaries.
  6. Schedule large batches overnight, as covered in Can a GB10 Run AI Agents 24/7 Without Anyone Watching?

Buying in Singapore

SourceIT supplies the ASUS Ascent GX10 (1 TB S$8,100, 2 TB S$8,300, 4 TB S$10,635), the NVIDIA DGX Spark (S$8,800) and other GB10 systems with local warranty and a GST invoice. Browse the NVIDIA DGX Spark collection or size a system with the AI hardware calculator. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

How long would a GB10 take to transcribe 100 hours of meetings? Based on published forum measurements, about four to five hours with Whisper large-v3 and under half an hour with NVIDIA Parakeet models, excluding speaker labelling and summaries.

Does Whisper work for Singapore meetings with mixed languages? Whisper is trained on many languages including Mandarin, Malay and Tamil, so it is the safer choice for multilingual meetings. Test it on your own recordings.

Does NVIDIA provide a transcription setup for DGX Spark? NVIDIA's DGX Spark playbooks do not include a speech-to-text recipe, but Whisper and Parakeet models run on the platform.

How much storage does 100 hours of audio take? About 11.5 GB as uncompressed 16 kHz mono audio, and much less compressed. The transcripts are only a few megabytes.

Can the same GB10 summarise the transcripts? Yes. A language model on the same system can produce summaries and action items, which should be checked by a person before they are relied on.

SourceIT Approved

SourceIT Approved

Researched and verified by the SourceIT team for Singapore buyers.

Share This Article
Facebook

More from our blog


← Older Post Newer Post →