Official provider results

Nemotron 3.5 Transcribe launch evaluation

NVIDIA reports 4.71% mean WER on the nine-set English Open ASR Leaderboard, 10.98% on the long-form board, 4.62% on the seven-language multilingual board and 15.80% mean cpWER on full-session multi-speaker recordings for Nemotron 3.5 Transcribe, plus 73.40% on MMAU and 66.67% on MMAR with reasoning on.

Source snapshots refreshed 2026-10-29

Five measurement lanesVOICE / S2S

Reasoning, task completion, conversation dynamics, experience, and latency stay separate.

Back to directory

Benchmark profile

Scope: English, long-form, multilingual and code-switched transcription, speaker-attributed transcription, contextual transcription with vocabulary hints, robustness, language identification, speech translation, audio understanding, spoken question answering, and text reasoning
Primary metric: Word, character and speaker-attributed error rates (%, lower is better); accuracy and BLEU (higher is better)
Owner: NVIDIA
Available evidence: Exact values in the Hugging Face model card tables

What it measures

Spoken reasoning

Does the system understand and answer spoken requests?

Voice experience

Is the exchange natural, robust, and responsive?

Available results

English and long-form transcription

NVIDIA-reported word error rate (%, lower is better). The nine-set mean is an equal-weight mean of the sets listed; the long-form mean averages CORAAL (itself averaged across locations), Earnings21 and Earnings22, processed in 55-minute chunks.

NVIDIA-reported word error rate (%, lower is better). The nine-set mean is an equal-weight mean of the sets listed; the long-form mean averages CORAAL (itself averaged across locations), Earnings21 and Earnings22, processed in 55-minute chunks.
Model	English 9-set mean	AMI	Earnings22	GigaSpeech	LibriSpeech clean	LibriSpeech other	Monsoon English	SPGISpeech	URGENT24	VoxPopuli	Long-form mean	CORAAL	Earnings21 (long)	Earnings22 (long)
Nemotron 3.5 Transcribe 30B A4B	4.71	7.81	5.29	7.18	1.09	2.60	3.65	1.91	11.00	1.84	10.98	13.39	7.57	11.99

Multilingual, code-switched and structured content

NVIDIA-reported error rates (%, lower is better). The multilingual board is the equal-weight mean of seven language means (Dutch, French, German, Hindi, Italian, Portuguese, Spanish) over FLEURS, Common Voice, Multilingual LibriSpeech and Monsoon; Hindi uses orthographically informed WER. ASCEND is a mixed error rate. HALAS is a robustness set.

NVIDIA-reported error rates (%, lower is better). The multilingual board is the equal-weight mean of seven language means (Dutch, French, German, Hindi, Italian, Portuguese, Spanish) over FLEURS, Common Voice, Multilingual LibriSpeech and Monsoon; Hindi uses orthographically informed WER. ASCEND is a mixed error rate. HALAS is a robustness set.
Model	Multilingual Open ASR	FLEURS English	ASCEND (Mandarin–English)	MUCS Hindi–English	MUCS Bengali–English	Monsoon Hindi (OIWER)	Numb3rs	VoiceCodeBench WER	HALAS
Nemotron 3.5 Transcribe 30B A4B	4.62	3.35	18.74	19.28	35.46	9.40	3.55	8.38	31.50

Multi-speaker and contextual transcription

NVIDIA-reported, lower is better. cpWER matches predicted speaker labels to reference speakers before scoring, on full recordings; cpCER is the character version for Mandarin. Contextual sets are scored without context and with the richest context NVIDIA reports (fine for ContextASR-Bench and Preference-ASR, local for Contextual Earnings22).

NVIDIA-reported, lower is better. cpWER matches predicted speaker labels to reference speakers before scoring, on full recordings; cpCER is the character version for Mandarin. Contextual sets are scored without context and with the richest context NVIDIA reports (fine for ContextASR-Bench and Preference-ASR, local for Contextual Earnings22).
Model	Mean cpWER (4 sets)	AMI IHM	AMI SDM	CALLHOME	Mixer 6	NOTSOFAR-1 IHM	NOTSOFAR-1 SC	AISHELL-4 cpCER	AliMeeting cpCER	ContextASR-Bench none / fine	Contextual Earnings22 none / local	Preference-ASR none / fine
Nemotron 3.5 Transcribe 30B A4B	15.80	10.29	17.34	11.50	14.80	14.94	25.91	32.91	34.95	3.33 / 1.71	12.54 / 9.58	7.56 / 6.16

Audio understanding and spoken instructions

NVIDIA-reported accuracy (%, higher is better) with reasoning on. MMAU-Pro rows are the non-music closed-form and instruction-following subsets; AIR-Bench is the foundation speech tasks. VoiceBench rows are individual subsets, not the leaderboard’s overall score. VoiceCodeBench TSR is the share of recordings with every target value correct.

NVIDIA-reported accuracy (%, higher is better) with reasoning on. MMAU-Pro rows are the non-music closed-form and instruction-following subsets; AIR-Bench is the foundation speech tasks. VoiceBench rows are individual subsets, not the leaderboard’s overall score. VoiceCodeBench TSR is the share of recordings with every target value correct.
Model	MMAR	MMAU	MMAU-Pro closed-form	MMAU-Pro instruction following	AIR-Bench speech	VoiceBench BBH	VoiceBench MMSU	VoiceBench OpenBookQA	VoiceBench IFEval	VoiceCodeBench TSR
Nemotron 3.5 Transcribe 30B A4B	66.67	73.40	52.69	86.21	67.27	84.40	79.28	91.21	81.47	59.33

Text reasoning

NVIDIA-reported accuracy (%, higher is better) on text inputs with reasoning on. These provider-run rows are shown for context and are not entered on BenchLM’s text benchmark lanes.

NVIDIA-reported accuracy (%, higher is better) on text inputs with reasoning on. These provider-run rows are shown for context and are not entered on BenchLM’s text benchmark lanes.
Model	GPQA	MMLU	MMLU-Pro	IFEval
Nemotron 3.5 Transcribe 30B A4B	62.12	83.99	74.14	84.65

Interpretation limit

NVIDIA ran every result itself. Transcription, translation and language identification use reasoning off at temperature 0; audio understanding and text reasoning use reasoning on at temperature 0.6, top-p 0.95, with a 16,384-token reasoning budget. The English Open ASR mean is NVIDIA’s equal-weight mean of nine sets, including Monsoon English and URGENT24, so compare it only with runs over the same set list. The multilingual board uses WER throughout, while standalone FLEURS uses CER for Chinese, Japanese, Korean, Thai and Vietnamese, so those numbers are not directly comparable. The card’s per-language FLEURS, CoVoST2, language-identification and translation tables are summarised, not copied in full. The “next best” comparisons in NVIDIA’s summary tables are not carried. These results do not enter text-model or voice-agent rankings.

Primary sources

Paper Code Owner page Hugging Face model card Technical report