NVIDIAAudio modelOpen weights
NVIDIA Nemotron 3.5 Transcribe
Summary
Nemotron 3.5 Transcribe is an open-weight speech model that combines audio recordings with text instructions. A parallel audio and speaker encoder feeds a Mamba-2 and Transformer mixture-of-experts decoder with 34 billion total parameters and about 4.3 billion active parameters. NVIDIA lists 40 transcription languages and recordings up to one hour long.
Depending on the instruction, it returns verbatim or normalized transcripts, speaker attribution, timestamps, translations, language identification, and answers to questions about audio. Its technical report spans Open ASR word error rates, speaker-attributed results, and contextual measurements. Quantization, benchmark setup, and hardware affect the scores, so the report separates BF16, FP8, NVFP4, and conditional DFlash2 tests.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 862 published benchmark results
Page 1 of 72
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
VoiceCodeBench, CTEM exact entity match (Show measurement, test conditions, and source)
- Source value
- 88.53
- Score
- 88.53
- Metric
- CTEM exact entity match
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 8
- Category
- entity-transcription
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; VoiceCodeBench English batch evaluation; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here. VoiceCodeBench batch mode; format-invariant WER.
VoiceCodeBench, TSR recordings with all entities correct (Show measurement, test conditions, and source)
- Source value
- 59.33
- Score
- 59.33
- Metric
- TSR recordings with all entities correct
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 8
- Category
- entity-transcription
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; VoiceCodeBench English batch evaluation; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here. VoiceCodeBench batch mode; format-invariant WER.
HALAS segments with insertions (Show measurement, test conditions, and source)
- Source value
- 40
- Score
- 40
- Metric
- Segments with insertions
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 8
- Category
- hallucination
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; HALAS Earnings-22 segments; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here. Automatic insertion measure on the test split.
Multilingual Open ASR Leaderboard, Hindi, public set, OIWER (Show measurement, test conditions, and source)
- Source value
- 9.15
- Score
- 9.15
- Metric
- OIWER
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 9
- Category
- multilingual-asr
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; Hindi, public set, OIWER; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here. Each language score averages FLEURS, Common Voice, and MLS where available; Hindi uses public Monsoon with orthographically informed WER. Mean is unweighted.
CS-FLEURS read, 14 language pairs, MER (Show measurement, test conditions, and source)
- Source value
- 31.06
- Score
- 31.06
- Metric
- MER
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 10
- Category
- code-switched-asr
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; CS-FLEURS read, 14 language pairs; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here.
CS-FLEURS MMS-TTS, 45 language pairs, MER (Show measurement, test conditions, and source)
- Source value
- 66.57
- Score
- 66.57
- Metric
- MER
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 10
- Category
- code-switched-asr
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; CS-FLEURS MMS-TTS, 45 language pairs; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here.
CS-FLEURS XTTS 1, 16 language pairs, MER (Show measurement, test conditions, and source)
- Source value
- 28.1
- Score
- 28.1
- Metric
- MER
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 10
- Category
- code-switched-asr
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; CS-FLEURS XTTS 1, 16 language pairs; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here.
CS-FLEURS XTTS 2, 60 language pairs, MER (Show measurement, test conditions, and source)
- Source value
- 41.68
- Score
- 41.68
- Metric
- MER
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 10
- Category
- code-switched-asr
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; CS-FLEURS XTTS 2, 60 language pairs; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here.
ASCEND Mandarin-English, MER (Show measurement, test conditions, and source)
- Source value
- 18.74
- Score
- 18.74
- Metric
- MER
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 10
- Category
- code-switched-asr
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; ASCEND Mandarin-English; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here.
ContextASR-Bench Speech English, no context, NE-WER (Show measurement, test conditions, and source)
- Source value
- 12.21
- Score
- 12.21
- Metric
- NE-WER
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 11
- Category
- contextual-asr
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; ContextASR-Bench Speech English, no context; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here.
ContextASR-Bench Speech English, no context, NE-FNR (Show measurement, test conditions, and source)
- Source value
- 14.99
- Score
- 14.99
- Metric
- NE-FNR
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 11
- Category
- contextual-asr
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; ContextASR-Bench Speech English, no context; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here.
ContextASR-Bench Speech English, coarse domain, NE-WER (Show measurement, test conditions, and source)
- Source value
- 12.26
- Score
- 12.26
- Metric
- NE-WER
- Unit
- %
- Benchmark version
- NVIDIA report 2026-10-28, Table 11
- Category
- contextual-asr
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- NVIDIA technical report
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. NVIDIA report; released Nemotron 3.5 Transcribe checkpoint after RL; ContextASR-Bench Speech English, coarse domain; BF16; default accuracy setup uses vLLM on one H100 GPU, MTP K=2, and benchmark-specific prompts/scoring. Exact task differences are stated here.
Page 1 of 72
All specifications
Architecture
- Architecture
- 34B total parameters including the audio encoder, about 4.3B active parameters; Mamba-2 and Transformer MoE with parallel audio and speaker encoders
- Source: Architecture
Input and output
- Languages
- 40 supported transcription languages; benchmark results cover different subsets by task
- Source: Languages
Licence and access
Further details
- Tasks
- Verbatim and written-form transcription, contextual and speaker-attributed transcription, rich transcription, speech translation, spoken language identification, and audio question answering
- Source: Tasks
- Checkpoint formats
- BF16, FP8, NVFP4, and GGUF. The optional DFlash2 draft is a separate model.
- Source: Checkpoint formats
- Evaluation setup
- Main results use the released BF16 checkpoint with vLLM and two MTP draft tokens on one H100 GPU. Reported values follow benchmark-specific datasets, prompts, and scoring protocols.
- Source: Evaluation setup
- Peak throughput
- Up to 2,238 audio seconds per wall-clock second with NVFP4 and DFlash2 on an RTX PRO 6000 Blackwell. The optimized NIM serving path is not part of the open vLLM integration shipped on release day.
- Source: Peak throughput
- Report evaluation checkpoint
- Main results use the released checkpoint after reinforcement learning. Transcription, translation, and language identification disable reasoning and use greedy decoding.
- Source: Report evaluation checkpoint
- English ASR protocol
- Section 4.2 uses the public leaderboard section with nine English test sets, cleaned current versions of AMI, Earnings-22, GigaSpeech, and VoxPopuli, and an unweighted mean.
- Source: English ASR protocol
- Speculative decoding protocol
- The report measures the released MTP module and the separate DFlash2 draft model for speculative decoding. DFlash2 is not the transcription checkpoint.
- Source: Speculative decoding protocol
- Unpublished statistical details
- The reviewed report tables provide scores and named test sets, but generally do not publish confidence intervals or per-test-set sample counts.
- Source: Unpublished statistical details
Sources and data date
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
| Evaluator | Test series | Composite indices | Recorded active measurements |
|---|---|---|---|
| Artificial Analysis | 3 | 1 | 10 |
Without vendor claims and historical or disputed results; those stay labelled in the single values. 6 measurements are not yet assigned to a test series.
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 29, 2026. 63 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | Accuracy-table confidence intervals Note The report does not publish confidence intervals for its accuracy tables. |
| Not published | Per-row benchmark sample counts Note Most accuracy rows do not state N. Exact sample counts are recorded where the report gives them; unpublished values remain unknown. |
| Not published | GPU configuration for MTP-head fine-tuning evaluation Note Appendix C.2 and Table 25 do not state the GPU configuration for this head-only fine-tuning comparison. The eight-H100 setup is explicitly reported for Appendix C.1/Table 24 and is not assumed here. |
| Not found in the inspected sources | Vals rows in captured payloads Note Nine complete Vals payloads dated 28 October were checked; no exact model ID or unambiguous model label. Limited to these saved benchmarks. |
| Unresolved | Uncaptured Arena Code/Speech leaderboards Note Text, Agent including Code/Chat/Work, and twelve WebDev views were captured. A separate general Code and Arena Speech leaderboard were not saved as complete tables. |
| Unresolved | VoiceArena exact Nemotron release row Note The current VoiceArena STT frontend exposes an NVIDIA / Nemotron 3.5 ASR row, but the captured material does not establish that it is the exact Nemotron 3.5 Transcribe 30B-A4B model. Its scores are intentionally not assigned. |
| Additional source | NVIDIA Nemotron 3.5 Transcribe project page (retrieved October 29, 2026) · NVIDIA Nemotron 3.5 Transcribe project page · Editorial description reviewed October 29, 2026 |
| Additional source | Nemotron 3.5 Transcribe technical report (retrieved October 29, 2026) · Nemotron 3.5 Transcribe technical report · Editorial description reviewed October 29, 2026 |
| Additional source | NVIDIA Nemotron 3.5 Transcribe model card (retrieved October 29, 2026) · NVIDIA Nemotron 3.5 Transcribe model card · Editorial description reviewed October 29, 2026 |
| Additional source | Artificial Analysis (retrieved October 29, 2026) |
| Additional source | Model metadata (retrieved October 29, 2026)1234 |