Nemotron Transcribe: check who said what
You have written notes from the planning meeting, but still have to replay the recording to find out who promised what. NVIDIA’s new Nemotron 3.5 Transcribe is worth considering if that happens often. Try it on an exchange where the speaker and the conditions matter. A tidy transcript that turns a suggestion into a promise creates more work.
What you can use now
TLDR AI’s edition dated 2026-10-29 led us to NVIDIA’s release on 2026-10-28. Nemotron 3.5 Transcribe can turn a recording into text, distinguish speakers, translate speech and answer questions about audio. You can guide the result with instructions, including vocabulary hints. We have checked the documentation, not tested the model.
NVIDIA makes the model weights available to download, so developers can run it themselves. Hosted services are another route. It is a component you can build into a process for handling recordings. It is not a finished app that joins meetings and distributes minutes.
The model works on recorded audio up to an hour long; it is not a live streaming transcription service. If you need captions while people are talking, this release does not meet that requirement.
Newsletter discovery: TLDR AI, 2026-10-29NVIDIA’s release and deployment options
The language count is only the first check
NVIDIA lists transcription in 40 languages, but speaker attribution in 13. Its model card includes English, French and German for both tasks. For Dutch and Polish, transcription is supported but multispeaker transcription is not specified. Check the task you need, as well as the language.
Speaker labels distinguish voices within a recording; they do not establish someone’s identity. You still need a reliable way to connect a label with a participant. Audio understanding and reasoning were evaluated only in English, so do not assume that support for a French or German transcript also guarantees reliable answers about it.
An illustrative example: the delivery that was never promised
This fictional exchange is a proposed test, not a customer story or a result we obtained. Two colleagues, labelled A and B, could record it expressly for the trial using invented details.
| Spoken exchange | What must survive in the notes |
|---|---|
| A: We could deliver on Friday. | Friday is a possibility, not a confirmed delivery. |
| B: Only if the stock arrives. Monday is confirmed. | The condition belongs to Friday; Monday is the confirmed date. |
| A: I will check the stock. I will not book the courier. | A will check stock. Nobody has agreed to book transport. |
Judge the transcript before asking for a summary
First ask for the words and speaker labels. Compare those with the recording. Then ask for a short list of commitments and unresolved questions. Keeping those two outputs separate helps you see whether a mistake began in the transcript or in the summary.
In the example, the summary should leave Friday conditional and assign the stock check to A. If it confirms Friday delivery or tells someone to book the courier, it fails the business task even if most words are correct.
Repeat the exercise with an interruption and the languages you use. Record material expressly for the trial. Count the time spent correcting speakers, dates and commitments, then compare the total effort with your current method.
Choose how much of the workflow you want to own
For occasional meetings, keep a tool that already produces useful notes. A model release alone is a weak reason to take on uploads, transcript storage, review screens and maintenance.
A hosted model can be worth exploring when you already have a repeatable recording workflow and a specific transcription problem. DeepInfra lists this exact model as a paid API and requires login for its demo. Confirm which model features its interface exposes, along with current pricing and processing terms, before planning an integration.
Running the model yourself gives you another deployment choice, with hardware and maintenance to budget for. NVIDIA’s reference setup uses an H100 80 GB GPU and documents other supported configurations. Downloadable weights do not make the complete service free.
Consider changing tools when the same errors repeatedly cost someone time and your comparison produces better notes. If the difficulty is agreeing responsibilities during the meeting, clearer decisions may help more than a new model.
DeepInfra: hosted model and current priceNVIDIA: deployment requirementsInclude review and maintenance in the cost
Sources
Start with one difficult exchange.
Describe the kind of recording that takes too long to turn into useful notes. forvertz can help plan a focused AI trial using an invented example. Do not share confidential client recordings.
Discuss a transcription workflowUpdated