Nemotron 3.5 Transcribe puts deployment claims to a local test

NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026, as an open-weight speech model for recorded audio. It combines transcription, translation, speaker attribution and natural-language control over the form of the output.

That combination may reduce the number of services in a meeting, interview or classroom workflow. It does not remove the need to test language coverage, hardware capacity, transcript conventions, human correction and audio governance together.

What the model does, and where the language count stops

NVIDIA says Nemotron 3.5 Transcribe covers 40 languages for transcription, translates 31 languages to or from English, and supports speaker-attributed transcription in 13 languages. The model accepts recordings of up to one hour and can return word-level timestamps linked to speakers.

Those are separate task counts, not one promise that every language supports every feature. The model card's language matrix should be checked against the intended combination of transcription, translation, diarization, script and code-switching. A language listed for transcription may not have the same support for translation or speaker attribution.

The model is a speech-augmented language model. Its speech encoder supplies audio information to a language model, while instructions select an output such as verbatim text, cleaned-up written text, contextual terminology or speaker labels. A prompt can provide likely names and domain terms, but it cannot prove that a name was spoken.

Speaker attribution divides an audio file into voices or turns. It does not authenticate the people in the room. If a transcript will be used to assign responsibility, record a decision or quote a participant, the recording and the label both need human review.

Why multilingual support is not a transcript acceptance test

Accent, microphone position, background noise, overlapping speech and specialist vocabulary create different error patterns. A clean single-speaker benchmark can therefore miss the failure that matters in a real meeting: a name changed into a common word, a sentence assigned to the wrong person or a timestamp shifted across an interruption.

Agora's published classroom demonstration shows one useful post-session workflow. A recorded class is sent to an early-access Nemotron model after the lesson, and the returned transcript with word-level timing is shown beside the replay. Agora describes the demo as a way to search a recorded lesson and jump to the relevant moment. It is a product demonstration, not evidence of live-caption latency or performance at the scale of a large course.

NVIDIA's medical speech case study gives a more demanding example. On held-out speakers in the Eka Care benchmark, the best prompt produced a word error rate of 22.44 percent and fine-tuning reduced it to 12.17 percent. Keyword word error rate fell from 9.2 percent to 6.5 percent. Those results belong to the published dataset, prompts and scoring pipeline; they do not predict the result for a hospital, call center or legal team with different names and microphones.

Code-switching needs its own sample. The HiACC study describes a Hinglish corpus containing adult and child recordings and reports that standard monolingual systems performed about 42 percent worse in word error rate on its test set. That finding concerns Hinglish and the study's baseline, not Nemotron's score. It does show why “multilingual” and “handles a language switch inside one sentence” must be tested separately.

Public leaderboards are useful for controlled comparisons. NVIDIA reports a mean word error rate of 4.71 percent on nine English Open ASR test sets and 4.62 percent on a multilingual leaderboard. The test sets, normalization rules and language mix still determine what those figures mean. An internal acceptance table should separately record word or character errors, entity errors, speaker swaps, timestamp drift, translation quality and correction time.

Self-hosting means checking more than GPU memory

NVIDIA publishes BF16, FP8, NVFP4 and GGUF checkpoints. The official project page lists vLLM, model-specific serving options and NeMo-Speech.cpp among the available paths, while the Hugging Face article also describes SGLang-Omni support. Each path has its own hardware, quantization and integration assumptions.

Active parameter count is not the same as total deployment memory. Measure the checkpoint, runtime, audio length, concurrency, queueing and retry behavior on the GPU that the team will actually operate. A one-hour recording processed after a meeting has a different service requirement from many short files arriving at once.

The model is released under the OpenMDW-1.1 license. “Open weight” describes access to model parameters; it does not make the software stack, training data, third-party packages and customer recordings share one license. Record the license owner and version for the model, inference server, tokenizer, data and any fine-tuning material before a production review.

Keeping audio in a company's own environment changes the transfer path, but it does not settle who may access the files, how long they remain, or whether a transcript is more widely visible than the original recording. Decide the purpose, access roles, retention period, deletion process, correction owner and failure fallback before connecting the model to meetings or customer calls.

How this plays out outside Taiwan

For an English-speaking team in the United States, Singapore, the United Kingdom or Australia, the first local question is not whether Nemotron can run offline. It is what the organization is allowed to do with the recording and transcript in the proposed workflow. A meeting file can contain names, voices, opinions, customer details, health information or confidential business plans. The deployment plan should name the purpose before the team chooses a checkpoint or writes a retention rule.

In the United States, there is no single nationwide rule that answers every workplace, school, health or customer-call recording question. The applicable answer can depend on the organization, the people recorded, the state and the purpose. A US team should identify where participants are located, check the relevant state recording-consent rules with counsel, and review sector requirements before recording. For a service used by children, the US Federal Trade Commission has treated voice recordings and transcripts as data that raise consumer privacy and security concerns. The FTC's discussion of Amazon's Alexa case says organizations should not retain children's voice data after the purpose has ended and should honor deletion obligations. That is a concrete reason to separate child-facing audio from ordinary internal meetings and to test whether backups, logs and derived transcripts are deleted together.

A US deployment should also record whether audio leaves the organization's network. “Self-hosted” may mean a GPU in a company facility, a managed cloud instance, or a hosted endpoint operated by another company. Those are different vendor and access arrangements. The procurement file should identify the processor, storage region, support access, incident process, model-improvement terms and deletion method. If the team cannot answer those questions from the vendor contract and privacy notice, it should keep the workflow in a pilot environment and use audio for which the organization has a clear basis.

Singapore gives a more direct official reference for recorded voices. The Personal Data Protection Commission says an audio recording may be personal data when a person can be identified from the voice or from the voice together with other information. The PDPC's guidance on audio recordings tells organizations to examine the recording in the context of identifiability. A Singapore team should therefore treat the original audio and the speaker-attributed transcript as records needing access control, not as harmless technical output. It should document the purpose, tell participants what is being recorded, restrict access to the people who need it, and set a deletion date that matches the purpose.

The same Singapore check applies when a team uses a local GPU but sends monitoring data, support bundles or failed requests to a third party. The data-flow diagram should include the audio, transcript, timestamps, speaker labels, prompts, error logs and any evaluation clips. If an evaluation set contains real customers or employees, it should be governed as carefully as the production set. If consent or notice does not cover model evaluation, use a controlled sample with the identifying material removed where that can be done without claiming that voice data has become anonymous.

In the United Kingdom, the Information Commissioner's Office says an organization should identify a particular need for audio recording, consider less intrusive alternatives and make the recording clear to people. The ICO's audio-recording guidance says audio should not be enabled by default where it is not justified and that people need additional notice. A UK employer considering Nemotron for meetings should document the purpose, consultation, access list, retention period and the less-intrusive options it considered. A transcript that identifies employees can remain personal data even when the original sound file is no longer in the main application.

The UK distinction between transcription and voice identification also matters. Transcribing a meeting and labeling turns as Speaker 1 and Speaker 2 is not the same operation as extracting voice features to identify a person against a stored voice profile. The ICO explains that voice recognition used for unique identification can involve biometric data, while an ordinary audio recording is not automatically biometric data merely because it contains a voice. The product design should state which operation is actually performed. Teams should not describe diarization as biometric identity verification, and should not enroll staff voices merely to make a transcript easier to read.

Australia's Office of the Australian Information Commissioner says a recorded voice can be personal information when the person is reasonably identifiable. The OAIC's explanation of personal information includes sounds and voice recordings, and points organizations to the Australian Privacy Principles. An Australian business should check whether the Privacy Act 1988 applies to it, whether a state or territory rule also applies, and whether the recording includes sensitive information. A health service provider or public body may have additional obligations, so “the audio stays on our server” is not a complete privacy assessment.

For all four markets, the practical local test is the same sequence with different authorities: list the people and countries involved; state the purpose; decide whether recording is necessary; provide the required notice or obtain the required permission; limit access; set retention and deletion; and test the transcript for names, speakers and sensitive terms. The official sources above explain how a jurisdiction treats recorded voices, but they do not certify Nemotron or approve a particular deployment. A team should save the source version, contract terms and decision record with its pilot results.

A small test can expose the deployment risk

Use audio that the organization is authorized to evaluate and create a human-checked reference transcript first. Compare the current workflow and Nemotron on the same files, with the same output rules. Include clean speech and difficult recordings, but keep the test set separate from production data until access and deletion controls are proven.

1. Define the task: separate verbatim text, cleaned-up text, translation, timestamps and speaker attribution. State how numbers, abbreviations, names and punctuation should appear.
2. Sample difficult audio: include accents, code-switching, specialist terms, overlapping speakers, distant microphones and background noise.
3. Build a human baseline: have someone who understands the context mark names, speaker turns and terms that must be preserved.
4. Measure the work: record word or character error rate, entity errors, speaker swaps, timestamp drift, GPU time, queue time, reruns and human correction minutes.
5. Set fallback levels: define what can be published after sampling, what needs full review and what returns to the old workflow.

The final comparison is the workflow, not the fastest model call. Nemotron 3.5 Transcribe is a candidate for recorded-audio processing when its language matrix, checkpoint, license and governance fit the use case. It needs stronger evidence before a team treats its output as a formal record, a live captioning service or proof of who said something.

Frequently asked questions

Does Nemotron 3.5 Transcribe support traditional Chinese?

NVIDIA's published matrix should be checked for the exact language and output task. A Mandarin transcription result does not by itself guarantee traditional-character formatting, regional vocabulary or mixed speech. Test representative recordings and inspect names and script conversion before adopting it.

Can it provide live meeting captions?

The published materials describe recorded-audio processing. Agora's classroom example sends audio after the class and displays a transcript with timing beside the replay. That is different from a tested live-streaming service with a stated latency and recovery target.

Do speaker labels identify the participants?

No. They distinguish voices or turns in the recording. They do not establish a speaker's real identity, and a person who knows the meeting still needs to verify any transcript used for decisions, accountability or formal records.