CekuraBench
Book a demo

NVIDIA Nemotron 3.5 Transcribe

Every modelNVIDIA

Public clips vs real calls

Lower and further left is better

Best quadrant
Ocular Dataset WER (%)
Pipecat Dataset WER (%)
Above the diagonal, a model does worse on real calls than on the public clips.

NVIDIA Nemotron 3.5 Transcribe is 2nd of 19 on both datasets together and 3rd of 19 on the Ocular Dataset. It is the only model on the board with open weights (OpenMDW-1.1).

2.28% Overall WER (2nd) · 1.89% Pipecat WER (2nd) · 3.03% Ocular WER (3rd)Likely range, 95% confidence: Pipecat WER 1.63 to 2.17%, Ocular WER 2.26 to 3.79%Open weights under OpenMDW-1.1, the only open-weights model on the board.Published price $0.12 per hour of audio. DeepInfra's hosted rate, $2 per 1,000 minutes. The scores come from a deployment hosted by NVIDIA, not DeepInfra's. Source, checked Oct 28, 2026.

Last updated Methodology by Dileep Chagam

NVIDIA Nemotron 3.5 Transcribe against every other model

Lower is better on every measure. Hover a column for that model's full numbers.

  1. 2.2%
    Transcribe Stream 1
  2. 2.3%
    Nemotron 3.5 Transcribe
  3. 2.4%
    Reson8
  4. 2.6%
    Ink 2
  5. 2.6%
    Realtime Whisper
  6. 2.7%
    Universal 3.6 Pro
  7. 2.7%
    Universal 3.5 Pro
  8. 2.8%
    Chirp 3
  9. 3.0%
    Linden
  10. 3.0%
    Chirp 2
  11. 3.0%
    STT-RT v5
  12. 3.6%
    Pulse
  13. 3.7%
    4o Mini Transcribe
  14. 3.8%
    Nova-3
  15. 4.0%
    Flux English
  16. 4.8%
    Flux Multilingual
  17. 5.4%
    STT-1
  18. 5.8%
    Gradium
  19. 6.9%
    4o Transcribe
Glossary: WER, TTFT, TTFS, P50
WER · Word error rate
Share of words the model got wrong: replaced, missing or added. Lower is better.
Pipecat Dataset
1,000 short public clips from Pipecat's STT benchmark dataset.
Ocular Dataset
8 recordings from 4 real phone conversations, licensed from Ocular.
Likely range · 95% range
Where the model's true score probably falls, given how much audio was tested. Overlapping ranges are too close to call.
TTFT · Time to first text
How long after the audio starts streaming until the first words appear.
TTFS · Final-text delay
How long after the caller stops until the full transcript arrives. A voice agent waits for this before it replies.
P50 / P90 / P95 · Typical, slow and slowest turns
Half of turns are faster than P50. Only 1 in 10 is slower than P90, and 1 in 20 slower than P95.

Where NVIDIA Nemotron 3.5 Transcribe struggled

On real calls, dropped words are its biggest source of error. On clean clips, wrong words lead.

37%

of its errors on the Ocular calls are words it left out.

Median across all 19 models: 49%

Mix of errors

Real calls (Ocular Dataset)

Clean clips (Pipecat Dataset)

TTFT and TTFS

Not measured. NVIDIA Nemotron 3.5 Transcribe was tested through an endpoint that takes a whole recording per request rather than a live stream, so there is no first text or final-text delay to time. Its turn transcripts are still scored for accuracy.

Every request used the verbatim transcription prompt from NVIDIA's model card, with reasoning off, and each 15-minute Ocular recording went in as a single request.

1,000 Pipecat clips · 8 Ocular recordings · 206 turns transcribed · 0 failed requests

NVIDIA Nemotron 3.5 Transcribe head to head

Every comparisonAll 19 modelsBenchmark source

NVIDIA Nemotron 3.5 Transcribe FAQ

  • NVIDIA Nemotron 3.5 Transcribe places highest on both datasets together, 2nd of 19 at 2.28%. Its weakest placing is the Ocular Dataset, 3rd of 19 at 3.03%.

  • NVIDIA Nemotron 3.5 Transcribe transcribes the 1,000 public Pipecat clips at 1.89% word error rate and the eight licensed Ocular recordings at 3.03%. Word error rate counts incorrect, missing, and extra words against the reference transcript, so lower is better.

  • NVIDIA Nemotron 3.5 Transcribe was tested through an endpoint that takes a complete recording per request, not a live stream, so there is no time to first text or final-text delay to measure. Its word error rates come from the same audio as every other model's.

  • Yes. Its weights are published under OpenMDW-1.1, so anyone can download the model and run it on their own hardware. It is the only open-weights model on the board; the other 18 are hosted APIs.

  • Every model runs the same two datasets under the same conditions: 1,000 public clips from Pipecat's STT benchmark dataset for accuracy at scale, and eight licensed conversation recordings from Ocular for real speech. NVIDIA Nemotron 3.5 Transcribe was sent each clip, recording and turn as a whole file, so it is scored on accuracy only. Every request used the verbatim transcription prompt from NVIDIA's model card, with reasoning off, and each 15-minute Ocular recording went in as a single request. NVIDIA Nemotron 3.5 Transcribe returned usable text for 1,000 of the public clips.

Building on NVIDIA? Test your agent the same way.

Cekura runs these simulated calls on your agent and scores every one.

Book a demoStart free