NVIDIA's Free Nemotron Transcribe AI Turns Meetings Into Text

NVIDIA open-sourced Nemotron 3.5 Transcribe, a free speech-to-text model that transcribes recordings in 40 languages, labels who is speaking, and adds timestamps on request. It tops NVIDIA's own Open ASR Leaderboards for accuracy among open-weights models, and the model weights are free to download and run.

What Is NVIDIA's Nemotron Transcribe?

Nemotron 3.5 Transcribe pairs a speech encoder with NVIDIA's Nemotron 3 diarization model and its Nemotron 3.5 Lightning language model, for about 4.3 billion active parameters working together. NVIDIA calls it a speech-augmented language model because a single model handles several jobs a transcription tool usually needs separate software for: a verbatim transcript, a cleaned-up written version, speaker-attributed text for a multi-person call, speech translation, and spoken language identification. The transcription style is chosen with a plain-language instruction for each request, with no retraining needed, and it works on recordings up to one hour long. NVIDIA says it is the first model of this kind to beat specialised, purpose-built systems at labelling who is speaking in overlapping, multi-person audio, which is normally the hardest part of transcribing a real meeting rather than a clean, one-person recording.

What Can It Actually Do?

• Transcribe speech in 40 languages, including code-switched speech where a speaker mixes languages mid-sentence
• Label which person is speaking, even in overlapping, multi-speaker recordings such as a group meeting or a busy call centre
• Add word-level, speaker-attributed timestamps for a recording on request
• Translate spoken content and identify which language is being spoken automatically
• Score a mean word error rate of 4.71 percent in English and 4.62 percent across its multilingual test set on NVIDIA's Open ASR Leaderboards, the best result of any open-weights model listed

How Could Your Business Use It?

Any business that records meetings, sales calls or customer service calls and never gets around to reviewing them is the obvious use case. A law firm could turn a client meeting recording into a written, timestamped record for the file. A customer service team could transcribe support calls automatically to spot recurring complaints without anyone listening back. A logistics company in Sabah handling calls that switch between English, Mandarin and Malay in the same conversation could get one written record instead of three separate transcripts, with each speaker's words correctly attributed to them. A training provider could turn a recorded workshop into searchable notes for attendees who missed a section. The catch is that this is a developer tool, not an app: a business needs someone who can set up the open source model and point it at a recording, not a button to click today.

Also Today

• Anthropic will switch Claude Enterprise to one unified chat-and-Cowork experience by default on 16 November 2026, so a conversation moves into a cloud sandbox automatically whenever a task needs files or code (Claude Help Center).
• Anthropic filed a registration statement with the US Securities and Exchange Commission on 26 October 2026 for a proposed initial public offering of its stock, a sign of how fast the company behind Claude is scaling (Anthropic).

What to Do This Week

If your team already records meetings or calls, ask whether anyone actually goes back and reviews them, and if the honest answer is no, that recording is wasted. You do not need to set up Nemotron Transcribe yourself to start building the habit: note down which recordings you would transcribe if it cost nothing, and bring that list into your next AI planning session. Our AI Tool Mastery for Work Productivity program covers exactly this kind of everyday automation, turning AI and Cowork into tools your team actually uses on real meetings and documents.