Nemotron 3.5 Transcribe

NVIDIA Nemotron 3.5 Transcribe (released as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B) is an open-weight speech and audio model from NVIDIA that turns recorded audio, together with a text instruction, into text. One checkpoint handles transcription in 40 languages, speaker-attributed transcription, word timestamps, contextual transcription with vocabulary hints, speech translation to and from English, spoken language identification, and audio question answering.^[1][2] NVIDIA released it on 28 October 2026 on Hugging Face, GitHub (through NeMo Speech), NGC and build.nvidia.com, under the OpenMDW-1.1 license.^[1][3]

The model is a speech-augmented language model (SALM): a speech recognition encoder and a frozen speaker diarization model feed audio embeddings into Nemotron 3.5 Lightning 30B-A3B, a hybrid Mamba-Transformer mixture-of-experts language model.^[2] It has about 34 billion parameters in total and about 4.3 billion active per token, including the audio encoder.^[1][2] NVIDIA describes it as an offline model for recorded audio, not a real-time streaming recognizer, and as a specialized transcription model rather than a general-purpose assistant.^[1][2]

Key facts

Release and availability

The model card lists the release date as 28 October 2026 for build.nvidia.com, GitHub (via NeMo Speech), Hugging Face and the NGC catalog.^[1] NVIDIA's research page and technical report are dated the same day.^[2][3] NVIDIA's Hugging Face blog post, written by members of the NVIDIA speech team, says the model is also available on ModelScope, with interactive demos on Hugging Face Spaces and ModelScope Studios, and through hosted endpoints from Baseten, DeepInfra and Simplismart.^[4] NVIDIA's announcement on X called it "an open-weight, promptable 30B MoE model with ~4B active parameters."^[5] The vLLM project announced "day-0 serving through vLLM + NeMo Speech" the same day.^[8]

NVIDIA released the training data pipeline alongside the model. Granary v2, the multilingual speech dataset and curation pipeline used to prepare much of the training data, was published on Hugging Face and ModelScope; NVIDIA's blog says its public manifest covers 463,062 hours across 48 languages and 16 source corpora, and that these figures describe the dataset release rather than the model's full training mixture.^[4]

The model card and technical report list separate FP8 and NVFP4 checkpoints, and NVIDIA's research page also lists a GGUF build for NeMo-Speech.cpp and a DFlash2 draft model.^[1][2][3] NVIDIA's Hugging Face blog post said on release day that "quantized versions are coming soon," and the FP8, NVFP4 and GGUF repository links returned an authentication error when checked on 29 October 2026, while the BF16 checkpoint and the DFlash2 draft model were public.^[4][1] Community conversions to GGUF, MLX and Core ML formats appeared on Hugging Face within a day of release.^[16]

Architecture

The design follows the SALM approach that NVIDIA researchers described in 2023: a speech encoder converts audio into embeddings, which are inserted into the prompt of a decoder-only large language model in place of an audio placeholder token, and the language model generates the response.^[2][15] The technical report names four departures from earlier SALMs: a Parallel Expert Encoder, a hybrid Mamba-Transformer MoE backbone, joint multi-token prediction, and a separate CTC adapter for word timestamps.^[2]

Parallel Expert Encoder

The encoder runs two branches over the same 128-bin log-mel features (16 kHz audio, 10 ms hop):^[2]

• ASR branch. A 32-layer Transformer encoder with model width 1,280, 10 attention heads, rotary position embeddings and query-key normalization. Frame stacking subsamples the features 8x, so each output frame covers 80 ms of audio. The branch has about 0.6B parameters and was initialized from a multilingual ASR encoder pretrained as a token-and-duration transducer (TDT) on Granary v2.^[2]
• Diarization branch. The Nemotron 3 Diarization model, a Sortformer speaker diarization model with about 100M parameters, estimates per-frame activity for up to eight speakers. The activities are binarized, normalized and projected through a fixed sinusoidal kernel into a speaker embedding that is added to the ASR representations. This branch stays frozen during training.^[2]

As the report describes the architecture, both branches process audio in independent 30-second chunks with no attention across chunk boundaries, and the encoder outputs are concatenated before they reach the language model, which is responsible for keeping speaker labels consistent across chunks.^[1][2] In the report's evaluation setup, the encoder instead uses 40-second chunks with 4 seconds of extra context on each side, and the diarization branch carries its speaker state from one chunk to the next.^[2] NVIDIA says the chunking is an internal processing choice and not a 30-second limit on recordings.^[1][2] A single linear projection maps the 1,280-dimensional encoder output to the 2,688-dimensional embedding space of the language model, so one second of audio becomes 12.5 language-model input positions.^[2]

Language model backbone

The backbone is Nemotron 3.5 Lightning 30B-A3B, with 52 layers: 23 Mamba-2 layers, 23 MoE feed-forward layers and 6 grouped-query attention layers. Each MoE layer has 128 routed experts with top-6 routing plus one shared expert, giving about 3.5B active parameters out of about 31.6B in the backbone.^[2] The report says the small number of attention layers keeps the key-value cache small, which matters for hour-long recordings.^[2] Adding the ASR encoder, diarization model, adapter and CTC head (about 0.8B parameters) gives about 4.3B active parameters, which is where the "A4B" in the name comes from; the MTP module (about 1.3B parameters) brings the checkpoint to about 34B in total.^[2]

Multi-token prediction and speculative decoding

The model was trained jointly with a multi-token prediction objective. The MTP module (one attention block and one MoE block) is shared across two prediction depths and reuses the backbone's output head; its loss is added to the main loss with weight 0.1.^[2] At inference the module acts as the draft model for speculative decoding in vLLM, and because it is shared between depths it can be applied more than twice to propose more draft tokens; the model card's recommended vLLM configuration (MTP8) proposes eight, and the report also evaluates 15.^[1][2] NVIDIA also trained a six-layer DFlash2 draft model as an alternative that proposes up to 15 tokens; MTP and DFlash2 are not combined, and word timestamps are not available with DFlash2.^[1][2]

Word timestamps

A connectionist temporal classification (CTC) adapter of about 50M parameters reuses the ASR encoder output, upsamples it to 40 ms frames and passes it through two Transformer layers. After the language model finishes generating, CTC forced alignment maps the transcript onto the audio, using the transcript's speaker tags and the diarization segments to produce speaker-attributed word timestamps. The adapter does not change the generated text.^[2] In vLLM, a request that enables timestamp capture receives a ctc_timestamps object with word start and end times, speaker tags, diarization segments and a mapping between transcript tags and diarization speakers.^[1]

Configuration summary

Capabilities and prompting

The task and output style are selected with a natural-language instruction in the user message; NVIDIA recommends leaving the system prompt empty for transcription and translation.^[1] The model card's prompt guide lists these uses:^[1]

Reasoning is controlled through the chat template (enable_thinking). NVIDIA recommends temperature 0 with reasoning off for transcription, translation and language identification, and temperature 0.6, top-p 0.95 with reasoning on for audio understanding.^[1] The training mixture also included small amounts of text instruction following and PII tagging and redaction tasks.^[1] The technical report says PII redaction and summarization have not been established as uniformly reliable and that model-based redaction should not be the only privacy control.^[2]

Languages

The model card's support matrix grades 40 transcription languages in three quality levels: "Leading" (state of the art or near it out of the box), "Production-ready", and "Supported" (may need fine-tuning).^[1] Translation covers both directions between English and the listed language. Audio understanding, reasoning and spoken instruction following were evaluated only in English.^[1]

Code-switching is marked as supported for all 40 languages.^[1] The matrix also contains a row for Norwegian with no ASR support marked; the technical report says the training data had 41 languages with at least 370 hours of distinct audio each and that 40 are supported because the Norwegian data "did not yield usable quality."^[1][2]

Training

Data

The model card reports about 1.72 million hours of audio across about 372 million unique indexed examples, or 6.53 million hours counting repeated exposure, plus 5.6 billion unique text tokens. Training used about 479 billion multimodal tokens in total, about 60% of them audio.^[1] The technical report says the speech data spans 77 languages, with English making up 36% of the audio seen in training.^[2] Transcription with and without context accounts for over 70% of training examples and speech translation for 17%.^[2]

Four sampling groups make up more than 95% of examples: Granary v2 pseudo-labeled data (31.2%), Granary v2 human-improved data (25.7%), internal speech data (25.1%) and Granary v2 speech translation (13.7%). Smaller sources teach specific skills such as multi-speaker transcription, audio understanding, rich transcription and PII tagging.^[2] Public sources named on the model card include LibriSpeech, Multilingual LibriSpeech, VoxPopuli, Common Voice 26, YODAS2, FLEURS, AMI, ICSI, AISHELL-4, AliMeeting, AudioSet, AudioCaps and MusicCaps; licensed or vendor data includes Fisher, Switchboard, Wall Street Journal, SPGISpeech and data from David AI.^[1] Multi-speaker training examples include constructed mixtures of up to eight speakers.^[1] NVIDIA states that user interactions with deployed products were not used for training.^[1]

Supervised fine-tuning and reinforcement learning

All components except the diarization model were trained jointly from the first step, without a separate modality-alignment stage.^[2] Supervised fine-tuning ran on 256 NVIDIA H100 GPUs (32 nodes of 8) with AdamW, a peak learning rate of 9 x 10^-5 and 65,536 packed tokens per GPU; the checkpoint after 28,500 steps was used for the next stage.^[2] A short reinforcement learning stage then applied Group Sequence Policy Optimization (GSPO) for 400 updates, with a reward of 1 minus the word error rate, on 20,000 selected examples (59% English).^[2] On NVIDIA's measurements, RL lowered English and multilingual error rates and improved number and code-entity transcription, and left translation and full-session multi-speaker results roughly unchanged. In the same matched comparison, MMAR audio-reasoning accuracy fell from 67.01% to 59.18%; the report notes that its later published scores differ because evaluation code and prompts were refined.^[2]

Benchmark results

Unless marked otherwise, the figures below are NVIDIA's own measurements from the model card (results captured on 12 October 2026) and the technical report.^[1][2] Error rates are percentages; lower is better.

English and long-form transcription

On the nine public test sets of the Hugging Face Open ASR Leaderboard, NVIDIA reports a mean word error rate of 4.71%. The report calls this the most accurate open-weight result on the public section of the leaderboard and says only seven commercial systems score better; its comparison table, which uses leaderboard baselines from 21 October 2026, lists six commercial systems, five of which score better than Nemotron 3.5 Transcribe.^[2]

Selected rows from Table 6 of the technical report.^[2]

Per-set results include 1.09% on LibriSpeech test-clean, 2.60% on test-other, 7.81% on AMI, 1.91% on SPGISpeech and 11.00% on the noisy URGENT 2024 set.^[1] On the long-form track (full earnings calls and interviews), NVIDIA reports a mean of 10.98%, against 13.37% for Parakeet-TDT-0.6B-v3 and 9.05% for ElevenLabs Scribe v2.^[2] The report notes that generation loops affected 7 of 451 long-form recordings and that the mean would be 12.87% without the repetition guardrail NVIDIA recommends applying to outputs.^[2]

Multilingual, code-switched and contextual transcription

On the multilingual Open ASR Leaderboard (German, French, Italian, Spanish, Portuguese, Dutch and Hindi), NVIDIA reports a mean of 4.62%, ahead of Qwen3-ASR-1.7B (6.49%) and Whisper-large-v3 (7.94%) and behind ElevenLabs Scribe v2 (3.56%) and Azure Speech (3.72%).^[2] Macro-averaged error rates are 9.29% on FLEURS (37 languages) and 12.19% on CoVoST 2 (17 languages).^[2] The report calls code-switching across many language pairs "a weakness of the model": it trails Whisper-large-v3 on the read and MMS-synthesized sets of CS-FLEURS.^[2]

Given a list of expected named entities on ContextASR-Bench (English), word error rate fell from 3.33% to 1.71% and named-entity WER from 12.21% to 2.20%.^[1][2] On Contextual Earnings-22, keyword recall rose from 58% to 96% with clip-level keywords while precision stayed near 98%.^[2] On the Preference-ASR style-control benchmark, the model scored 7.56% with a generic prompt and 6.16% with the style instruction, against 8.16% and 8.36% for Canary-Qwen-2.5B.^[1][2]

Multi-speaker transcription

NVIDIA's strongest claims concern speaker-attributed transcription of overlapping speech. Scored with cpWER (word error rate after matching predicted speakers to reference speakers) on full sessions, NVIDIA reports these results against open-weight systems it ran with the same data and scoring:^[2]

The report says that, to the authors' knowledge, this is the first SALM to outperform specialized cascaded systems on this task, though DiCoW and MOSS-Transcribe-Diarize are more accurate on some individual sets.^[2] Mandarin far-field meetings remain weak (cpCER of 32.91% on AISHELL-4 and 34.95% on AliMeeting), which the report attributes to very little Mandarin multi-speaker training data.^[2] Speaker attribution supports at most eight speakers, and NVIDIA notes that speaker labels distinguish voices within a recording and do not establish real-world identity.^[1][2]

NVIDIA's blog also cites third-party speaker benchmarks: first place among 14 systems on VoiceArena's English Diarization-Bench (32.4% cpWER), and the lowest diarization error rate (18.26%) of 14 systems on David AI's 21-language DAI-ASR-I18N diarization track, slightly ahead of the standalone Nemotron 3 Diarization model.^[4] These figures are as reported by NVIDIA.

Translation, language identification and audio understanding

Averaged over languages, NVIDIA reports 29.20 BLEU into English and 26.54 BLEU out of English on FLEURS, and 33.67 and 26.34 BLEU on CoVoST 2.^[2] Language identification accuracy averages 98.64% over 36 FLEURS languages and 95.58% over 24 VoxLingua107 languages, with the weakest results for Latvian and Slovenian.^[2]

With reasoning enabled, the model scores 73.40% on MMAU, 66.67% on MMAR (speech subset) and 83.8 overall on VoiceBench.^[1][2] The technical report says it is ahead of smaller open audio models on several of these tests but behind Qwen3-Omni-30B-A3B and the best published results, and that its voice instruction following trails NVIDIA's general-purpose Nemotron 3 Nano Omni.^[2] Text-only results (with reasoning) were 62.12% on GPQA Diamond and 74.14% on MMLU-Pro, compared with 75.44% and 81.94% reported for the Nemotron 3.5 Lightning base model; NVIDIA says preserving the base model's text abilities was "explicitly not a goal."^[2]

Independent leaderboards

When checked on 29 October 2026, the Hugging Face Open ASR Leaderboard's open-weight view ranked the model first with an average WER of 5.33% and an RTFx of 462.09. That default ranking adds two private test sets ("scripted" and "conversational") to the nine public ones, which is why it differs from NVIDIA's 4.71% public-set figure. Qwen3-ASR-1.7B was second at 5.40%.^[9] The leaderboard's changelog lists the model as added on 28 October 2026.^[9] NVIDIA's launch thread cited the same 5.33% figure.^[6]

Artificial Analysis lists the model on its non-streaming speech-to-text leaderboard. On 29 October 2026 its AA-WER v2 index chart (weighted 50% AA-AgentTalk, 25% VoxPopuli-Cleaned-AA, 25% Earnings22-Cleaned-AA) showed Nemotron 3.5 Transcribe at 2.8%, level with Mistral's Voxtral Small.^[10] The provider table on the same page listed the DeepInfra-hosted endpoint at 3.4% WER, a median speed factor of 73.8 and a price of $2.00 per 1,000 minutes of audio.^[10] NVIDIA's thread put the figures at 2.81% for Nemotron 3.5 Transcribe and 2.77% for Voxtral Small, noting the roughly sixfold difference in active parameters (about 4B against 24B).^[7] The Decoder reported the same comparison.^[11]

Quantization and inference efficiency

According to the technical report, the FP8 and NVFP4 checkpoints are produced by post-training quantization with NVIDIA Model Optimizer. FP8 halves the checkpoint from 67.5 GB to about 35 GB; the mixed-precision NVFP4 checkpoint, intended for Blackwell GPUs, is about 21 GB.^[2] In NVIDIA's comparison on one GB300 GPU, English Open ASR accuracy was nearly unchanged (4.70% BF16, 4.74% FP8, 4.73% NVFP4), while NVFP4 cost more accuracy on multilingual and distant-microphone tests; NVIDIA recommends FP8 where non-English accuracy matters and NVFP4 where memory or throughput is the main constraint.^[2]

NVIDIA measured throughput with its Nemotron Transcribe NIM, which uses serving optimizations not yet in the open-source release. On one RTX PRO 6000 Blackwell Server Edition GPU, NVFP4 with DFlash2 reached 2,238 seconds of audio per second and returned a 30-second transcript in 0.13 seconds for a single request; with MTP, throughput was 1,021 RTFx in BF16, 1,530 in FP8 and 1,811 in NVFP4 on 30-second clips.^[1][2] NVIDIA says the vLLM integration shipped in NeMo Speech on release day is intended for model development and is expected to be roughly two times slower.^[1] For batched English transcription on one H100 with the open release, the model card gives approximate BF16 throughput of 410 RTFx with MTP2 and 340 with DFlash2 (K=7).^[1]

Deployment

The model card lists one H100 80 GB as the reference deployment.^[1] Supported options on release:

The hardware table covers Ampere (A100, BF16), Hopper (H100 and H200, BF16 and FP8), Ada Lovelace (L40S, FP8) and Blackwell (B200, GB200, DGX Spark, GB300 DGX Station, Jetson Thor and RTX PRO 6000 in all three precisions; RTX 5090 in NVFP4 only).^[1] Fine-tuning tools ship in NeMo Speech. The technical report says the model was adapted to 17 tasks on a single GB300 GPU, typically in 0.5 to 2.3 GPU-hours per task.^[2] In a medical case study on Eka Care's Indian medical speech benchmark, NVIDIA reports that fine-tuning (with LoRA on the language model) cut held-out word error rate from 22.44% with the best prompt to 12.17%.^[4]

License

The weights are released under the OpenMDW License Agreement, version 1.1, a permissive license published by the Linux Foundation for model materials.^[1][12] It grants permission to use, modify and distribute the model materials, requires redistributors to keep the agreement and notices, terminates the rights of anyone who files a lawsuit claiming the materials infringe a patent or copyright (unless responding to a suit brought against them first), and places no restrictions on outputs.^[12] The model card notes that some included code, such as a vLLM patch and a repetition-loop guardrail adapted from Qwen3-ASR, is under the Apache License 2.0 instead.^[1] The card states the model "is ready for commercial or non-commercial use."^[1]

Limitations

The technical report and model card list these limitations:^[1][2]

• It is an offline model; internal 30-second chunking does not make it a streaming recognizer, and streaming support is described as future work.
• Behavior beyond one hour of audio or 65,536 tokens is not guaranteed.
• Audio understanding and reasoning were evaluated only in English.
• Code-switching across some language pairs, including Chinese-English, and Mandarin far-field meetings remain weak.
• The model can produce spurious transcripts for non-speech audio, return empty outputs or refusals for valid requests, and fall into generation loops; NVIDIA supplies a text-level guardrail for repetition loops.
• Output conventions such as number formatting and script depend on the wording of the instruction.

Relation to other NVIDIA speech models

Nemotron 3.5 Transcribe sits beside NVIDIA's earlier open ASR models, the Parakeet transducer and CTC models and the Canary encoder-decoder models, which the technical report groups with Whisper among the attention encoder-decoder and transducer models that drove recent progress in speech recognition.^[2] Canary-Qwen-2.5B, an earlier NVIDIA SALM, and Parakeet-TDT-0.6B-v3 appear as baselines in its English, long-form and style-control comparisons.^[2] The model reuses Nemotron 3 Diarization as a frozen component and takes its language backbone from the Nemotron 3.5 Lightning LLM.^[2] It is distinct from Nemotron 3.5 ASR, a 600M-parameter cache-aware FastConformer-RNNT model for streaming transcription in 40 language-locales that NVIDIA released on Hugging Face on 4 June 2026.^[17] It is distributed through NVIDIA NeMo Speech, where the training recipe and fine-tuning toolkit are published.^[1][2] Other 2026 transcription products covered on this wiki include GPT-Transcribe, Cohere Transcribe, Gemini 3.5 Transcribe and Muse Voice Transcribe.

Continue reading

• NvidiaNvidia Corporation is an American technology company that designs graphics processing units (GPUs) and is the world's dominant supplier of hardware for artificial intelligenceMentioned 55 times above
• NemotronNemotron is NVIDIA's brand for its family of open large language models and the datasets, training recipes, and evaluation tools built around them.Mentioned 24 times above
• vLLMvLLM is an open-source inference and serving engine for large language models. It is software for running already trained models, either through a Python interface for offline work or through a network server…Mentioned 8 times above
• NNemotron 3.5 LightningNVIDIA Nemotron 3.5 Lightning is an open-weights 30 billion parameter mixture-of-experts language model with 3 billion active parameters per token, released by NVIDIA on August 11Mentioned 6 times above
• NNVIDIA ParakeetParakeet is a family of open automatic speech recognition (ASR) models developed by NVIDIA as part of the NeMo conversational AI toolkit.Mentioned 5 times above
• NVIDIA CanaryCanary is a family of open speech models developed by Nvidia as part of its NeMo conversational AI toolkit.Mentioned 4 times above

More to read

• Speech Recognition
• NVIDIA NeMo
• NVFP4
• Word error rate
• GPT-Transcribe
• Cohere Transcribe

Topics:AI ModelsNVIDIAOpen Source AISpeech & Audio AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53NVIDIA, "NVIDIA-Nemotron-3.5-Transcribe-30B-A4B" model card, Hugging Face. Retrieved 29 October 2026. huggingface.co/...-Nemotron-3.5-Transcribe-30B-A4B
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54 ^55 ^56 ^57 ^58 ^59 ^60 ^61 ^62 ^63 ^64 ^65 ^66 ^67 ^68 ^69 ^70 ^71 ^72 ^73NVIDIA, "Nemotron 3.5 Transcribe Technical Report," 28 October 2026. research.nvidia.com/...scribe-Technical-Report.pdf
3. ^1 ^2 ^3 ^4NVIDIA Research, "NVIDIA Nemotron 3.5 Transcribe: Intelligent Transcription with a Speech LLM," published 28 October 2026. research.nvidia.com/...DIA-Nemotron-3.5-Transcribe
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Maryam Motamedi, Piotr Żelasko, Kunal Dhawan et al. (NVIDIA), "NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt," Hugging Face blog, 28 October 2026. huggingface.co/...nemotron-transcribe-granary
5. ^NVIDIA AI (@NVIDIAAI), "Meet NVIDIA Nemotron 3.5 Transcribe," X, 28 October 2026. x.com/...2115462300241141882
6. ^NVIDIA AI (@NVIDIAAI), thread post 2/ on the Open ASR Leaderboard ranking, X, 28 October 2026. x.com/...2115462302078320816
7. ^NVIDIA AI (@NVIDIAAI), thread post 5/ on Artificial Analysis results, X, 28 October 2026. x.com/...2115462308445262329
8. ^vLLM (@vllm_project), "NVIDIA Nemotron 3.5 Transcribe is here," X, 28 October 2026. x.com/...2115465359298077164
9. ^1 ^2 ^3Hugging Face, Open ASR Leaderboard (open-weight view and changelog). Retrieved 29 October 2026. huggingface.co/...open_asr_leaderboard
10. ^1 ^2Artificial Analysis, "Speech to Text AI Model & Provider Leaderboard" (non-streaming). Retrieved 29 October 2026. artificialanalysis.ai/...non-streaming
11. ^Jonathan Kemper, "Nvidia's open Nemotron 3.5 Transcribe combines transcription, translation, and speaker recognition," The Decoder, 28 October 2026. the-decoder.com/...slation-and-speaker-recognition
12. ^1 ^2OpenMDW, "OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)." Retrieved 29 October 2026. openmdw.ai/...1-1
13. ^Hugging Face, "Nemotron 3.5 Transcribe," Transformers documentation (main). Retrieved 29 October 2026. huggingface.co/...nemotron3_5_transcribe
14. ^SGLang-Omni, "Nemotron Transcribe" cookbook. Retrieved 29 October 2026. sgl-project.github.io/...nemotron_transcribe
15. ^Zhehuai Chen, He Huang, Andrei Andrusenko, Oleksii Hrinchuk, Krishna C. Puvvada et al., "SALM: Speech-augmented Language Model with In-context Learning for Speech Recognition and Translation," arXiv:2310.09424, 2023. arxiv.org/...2310.09424
16. ^Hugging Face model search for "Nemotron-3.5-Transcribe." Retrieved 29 October 2026. huggingface.co/models
17. ^NVIDIA, "nemotron-3.5-asr-streaming-0.6b" (Nemotron 3.5 ASR) model card, Hugging Face, released 4 June 2026. Retrieved 29 October 2026. huggingface.co/...nemotron-3.5-asr-streaming-0.6b

Show all 17 references