· Oct 28, 2026
NVIDIA presents Nemotron 3.5 Transcribe and Granary v2 for adaptable speech recognition
Markdown · JSON · Original source
’s article presents Nemotron 3.5 Transcribe, an for transcription, speaker attribution, translation and audio question answering, alongside Granary v2, a multilingual data curation pipeline. The article says developers can guide output with natural-language prompts and adapt the model to domain vocabulary and accents using tools. NVIDIA reports strong results on several speech , but those scores measure performance on specific , not reliability in every application. In an Indian medical speech case study, the authors report that fine-tuning reduced word error rate on held-out speakers; general English word error rate rose slightly, and a medical specialist still led on exact matches for brand-name drugs and diagnostics. Granary v2’s public manifest covers 463,062 hours across 48 languages, but those figures describe the dataset release, not the model’s full training mixture or sampling weights. The article also cautions that task support differs by language, audio question answering is trained in English, and the model is built for recorded audio rather than live streaming. How well would the reported medical gains hold up across other clinical settings and speech conditions?
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Oct 28, 2026 (2026-10-28T14:58:17.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-14T07:46:47.083Z. First observed: 2026-10-28T14:58:33.792Z. Retrieved: 2026-10-28T15:00:39.519Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The performance claims come from and a medical case study; the article does not establish results across all real-world applications or clinical settings. The authors report a slight increase in LibriSpeech test-clean word error rate after adaptation and say a medical specialist still leads on exact matches for brand-name drugs and diagnostics. The article says task support differs by language and audio question answering is trained in English. The model is built for recorded audio, not live streaming. The public manifest’s hours and language count do not describe the model’s full training mixture or sampling weights.
Open question
How well would the reported medical gains hold up across other clinical settings and speech conditions?
Source-stated
NVIDIA describes Nemotron 3.5 Transcribe as an open-weight model that combines transcription, speaker attribution, translation and audio question answering.
It brings transcription, speaker attribution (diarization), translation, and audio question answering into one checkpoint
Source-stated
The article says developers can prompt the model and adapt recognition to domain vocabulary and accents.
Developers can guide output with natural language, choose where their audio is processed, and adapt recognition to domain vocabulary and accents.
Source-stated
NVIDIA reports that fine-tuning reduced held-out word error rate in its medical speech case study, while general English word error rate rose slightly.
fine-tuning cuts held-out WER from 22.44% for the best-prompt baseline to 12.17%, a 45.8% relative reduction
Source-stated
The article says Granary v2’s public manifest covers 463,062 hours across 48 languages, and clarifies that these figures do not describe the model’s full training mixture or sampling weights.
The public manifest covers 463,062 hours across 48 languages and 16 source corpora
Source-stated
NVIDIA says the model is designed for recorded audio rather than live streaming, and that support varies by task and language.
language support differs by task, and audio question answering is trained in English.
Source and contributor context
Hugging Face · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Open weights
A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Held-out set
Data kept aside during training and used only to test how well a model works on examples it has not seen.
NVIDIA
NVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Editorial provenance
Version 1. Publication ID: caa5df6e-1cbb-4d53-8b4e-832d433bfd13.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-14T07:46:47.083Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b1607fbfd04a60705a33fa4d081c9214fcc5e68aa55d622506c889af20c657b1. Source text hash: c4eeb2b6d7828b46e12d25d0c6910f3ef92779083f4f4f7160b8fa47abdc9bbf.
Revision and source-check history
Version 1 · 2026-11-14T07:46:47.083Z
NVIDIA presents Nemotron 3.5 Transcribe and Granary v2 for adaptable speech recognition
’s article presents Nemotron 3.5 Transcribe, an for transcription, speaker attribution, translation and audio question answering, alongside Granary v2, a multilingual data curation pipeline. The article says developers can guide output with natural-language prompts and adapt the model to domain vocabulary and accents using tools. NVIDIA reports strong results on several speech , but those scores measure performance on specific , not reliability in every application. In an Indian medical speech case study, the authors report that fine-tuning reduced word error rate on held-out speakers; general English word error rate rose slightly, and a medical specialist still led on exact matches for brand-name drugs and diagnostics. Granary v2’s public manifest covers 463,062 hours across 48 languages, but those figures describe the dataset release, not the model’s full training mixture or sampling weights. The article also cautions that task support differs by language, audio question answering is trained in English, and the model is built for recorded audio rather than live streaming. How well would the reported medical gains hold up across other clinical settings and speech conditions?
Original source
· Article
Source details
Original source date: Oct 28, 2026 (2026-10-28T14:58:17.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-14T07:46:47.083Z. First observed: 2026-10-28T14:58:33.792Z. Retrieved: 2026-10-28T15:00:39.519Z.
Limits
The performance claims come from and a medical case study; the article does not establish results across all real-world applications or clinical settings. The authors report a slight increase in LibriSpeech test-clean word error rate after adaptation and say a medical specialist still leads on exact matches for brand-name drugs and diagnostics. The article says task support differs by language and audio question answering is trained in English. The model is built for recorded audio, not live streaming. The public manifest’s hours and language count do not describe the model’s full training mixture or sampling weights.
Open question
How well would the reported medical gains hold up across other clinical settings and speech conditions?
Source-stated
NVIDIA describes Nemotron 3.5 Transcribe as an open-weight model that combines transcription, speaker attribution, translation and audio question answering.
It brings transcription, speaker attribution (diarization), translation, and audio question answering into one checkpoint
Source-stated
The article says developers can prompt the model and adapt recognition to domain vocabulary and accents.
Developers can guide output with natural language, choose where their audio is processed, and adapt recognition to domain vocabulary and accents.
Source-stated
NVIDIA reports that fine-tuning reduced held-out word error rate in its medical speech case study, while general English word error rate rose slightly.
fine-tuning cuts held-out WER from 22.44% for the best-prompt baseline to 12.17%, a 45.8% relative reduction
Source-stated
The article says Granary v2’s public manifest covers 463,062 hours across 48 languages, and clarifies that these figures do not describe the model’s full training mixture or sampling weights.
The public manifest covers 463,062 hours across 48 languages and 16 source corpora
Source-stated
NVIDIA says the model is designed for recorded audio rather than live streaming, and that support varies by task and language.
language support differs by task, and audio question answering is trained in English.
Source and contributor context
Hugging Face · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Open weights
A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Held-out set
Data kept aside during training and used only to test how well a model works on examples it has not seen.
NVIDIA
NVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Editorial provenance
Version 1. Publication ID: caa5df6e-1cbb-4d53-8b4e-832d433bfd13.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-14T07:46:47.083Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b1607fbfd04a60705a33fa4d081c9214fcc5e68aa55d622506c889af20c657b1. Source text hash: c4eeb2b6d7828b46e12d25d0c6910f3ef92779083f4f4f7160b8fa47abdc9bbf.
passed · 2026-11-14T07:46:47.083Z · anthropic/claude-opus-5.5
Publication: caa5df6e-1cbb-4d53-8b4e-832d433bfd13
attribution — pass: The reading attributes the model, benchmark results and case-study findings to NVIDIA and the article's authors. It presents them as reported claims rather than established facts. The byline lists NVIDIA authors on a Hugging Face blog post.
NVIDIA Nemotron 3.5 Transcribe is an open-weight, promptable model that turns recorded speech into useful text
summary — pass: The reading names the actor and the release, and explains in plain language what the model does. It keeps the source's main caveats: the benchmark scope, the slight rise in LibriSpeech error, the specialist model's lead on brand-name drugs, the dataset-versus-training-mixture distinction, English-only audio Q&A and recorded-audio-only use. It is somewhat generic on the leaderboard results and omits the 45.8% figure from the summary text, but these are stylistic choices. The question follows from the case study's limits.
On general English, LibriSpeech test-clean WER rises slightly, from 1.54% to 1.69%.
date — pass: The originalPublished date of 2026-10-28 matches the page's own publication line of October 28, 2026.
Published October 28, 2026
support — pass: Every claim is source-stated. The held-out reduction, the LibriSpeech rise, the specialist's brand-name lead, the manifest figures and their caveat, and the task and language limits all appear in the source. The recorded-audio limit is also explicitly stated in the source. The reading does not overstate benchmark scores as real-world reliability.
Transcribe is built for recorded audio, not live streaming.
media — pass: The selected media is the verified direct link to the original article, which is the most direct way to reach the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceNVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Explore NVIDIAA model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceData kept aside during training and used only to test how well a model works on examples it has not seen.
Definition referenceHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceNVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Explore NVIDIAA model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceData kept aside during training and used only to test how well a model works on examples it has not seen.
Definition referenceHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference