NVIDIA Releases Nemotron 3.5 Transcribe Open-Weights STT Model

NVIDIA Releases Nemotron 3.5 Transcribe Open-Weights STT Model

NVIDIA dropped Nemotron 3.5 Transcribe on October 28 — a 30B-parameter (4B active) open-weights speech LLM that tops the open-model leaderboards: 2.8% WER on Artificial Analysis, 5.33% on the HF English Open ASR leaderboard, and 4.62% multilingual [1][2]. It handles verbatim, contextual, and speaker-attributed transcription, translation across 40 languages, diarization for up to eight speakers, and even audio Q&A — all in a single pass on recordings up to an hour long [3].

The performance numbers are notable less for raw accuracy and more for speed and versatility: up to 2,238 seconds of audio processed per second on a Blackwell GPU, plus a strong 1.8% score on the AA-AgentTalk voice-agent benchmark [2]. Weights are free on Hugging Face and ModelScope.

Developers on X zeroed in on its handling of overlapping speech and multi-speaker meetings — exactly the messy real-world audio that trips up lesser ASR systems. For any vendor building transcription infrastructure, an open model this capable raises the bar on what "good enough" diarization and accuracy now mean.

Graphify Converts Folders of Code, Docs, and Images into Queryable Personal Knowledge Graphs

Graphify, an open-source local tool, turns folders of code, documents, and images into interactive, queryable knowledge graphs — no cloud, no vector database required [1][2]. It uses tree-sitter to deterministically parse 36+ programming languages plus Markdown and SQL into AST-based nodes and relationships, layering semantic extraction on top for docs, PDFs, and images [3]. Output includes an interactive graph.html, a GRAPH_REPORT.md, and a graph.json — each relationship tagged as EXTRACTED or INFERRED for provenance.

It plugs into Obsidian and Claude Code, and can be invoked via CLI or a simple /graphify . slash command inside an AI assistant.

X discussion centered on second-brain systems — builders see Graphify as a step beyond flat semantic search toward structured, explainable retrieval (shortest paths, traceable reasoning) across your own files. It's a small project, but it signals where personal knowledge tooling is heading: local-first, structured, and provenance-aware.

Granola Meeting Transcripts Integrated into Obsidian PKM and Agent Workflows

Community plugins like obsidian-granola-plugin and og-sync now pull Granola's AI meeting summaries, notes, and full transcripts — via the official MCP API/OAuth or manual exports — directly into Obsidian vaults as structured Markdown, complete with frontmatter for attendees, timestamps, and cross-links [1][2]. The plugins handle deduplication and attendee linking, and feed into agent pipelines that do summarization, todo extraction, and routing to downstream tools like AssemblyAI or Dropbox [3].

This is grassroots infrastructure-building — users and developers stitching together a tool that only captures meetings (Granola) with a tool that's built for durable personal knowledge (Obsidian), because neither one does the full job alone.

X conversation focused squarely on the pipeline angle: automated, agent-driven workflows that make meeting transcripts actually usable weeks or months later, rather than sitting in a silo. It's a trend worth watching — people are hungry for tools that treat meetings as a knowledge source, not a disposable recording.

What This Means For Your Meetings

Today's news traces a single arc: the meeting is no longer an isolated event — it's one node in an ongoing, queryable record of your work. Otter's Work Mode explicitly reaches backward into your meeting history and forward into your CRM; Granola users are hand-building that same bridge into Obsidian because the market hasn't fully delivered it yet; and Graphify shows the underlying appetite for structured, explainable knowledge graphs over flat note dumps. The common thread is that transcription itself is becoming table stakes — NVIDIA's Nemotron 3.5 Transcribe, free and open-weights, all but guarantees that raw speech-to-text accuracy stops being a differentiator within the year.

That's exactly the terrain Proudfrog is built for. When transcription is commoditized, the value shifts entirely to what happens after the recording stops: how speakers are identified and tracked over time, how a knowledge graph connects today's decision to last quarter's discussion, and how AI retrieval surfaces the right fact from the right meeting months later without you remembering which call it was in. The DIY pipelines people are assembling with Granola and Obsidian today are a preview of what needs to be native, not bolted-on.

The lesson for meeting-intelligence vendors is clear: briefs and recaps are necessary but no longer sufficient. The real competitive ground is durable, structured memory across your entire meeting history — connected, provenance-aware, and instantly retrievable.

Key takeaway: As transcription accuracy becomes free and commoditized, the battle moves to memory — who can turn a year of meetings into a trustworthy, queryable knowledge base, not just a stack of transcripts.

Sources

1. https://otter.ai/blog/otter-rethinks-the-meeting-new-ai-preps-you-coaches-you-and-follows-through
2. https://www.computerworld.com/article/4241172/otter-adds-work-mode-to-automate-more-meeting-related-tasks.html
3. https://otter.ai/work-mode
4. https://research.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Transcribe/
5. https://siliconangle.com/2026/10/28/nvidia-releases-all-in-one-adaptable-nemotron-3-5-audio-transcription-and-translation-model/
6. https://huggingface.co/blog/nvidia/nemotron-transcribe-granary
7. https://graphify.com/docs/tutorial
8. https://github.com/Graphify-Labs/graphify
9. https://graphify.net/knowledge-graph-for-ai-coding-assistants.html
10. https://github.com/philfreo/obsidian-granola-plugin
11. https://github.com/og-sync/obsidian-granola-sync
12. https://superframeworks.com/articles/granola-encrypted-local-notes