Smallest.ai Ships Pulse STT With Sharper Multi-Speaker Diarization

Smallest.ai Ships Pulse STT With Sharper Multi-Speaker Diarization

Smallest.ai's Pulse model now handles speaker diarization in both batch and streaming modes, assigning confidence-scored speaker labels in real time [1]. The headline spec is speed — 64ms latency across 38+ languages — but the more interesting detail for meeting-heavy teams is the speaker-count ceiling: non-streaming mode caps reliably at four speakers, with accuracy degrading beyond that, while streaming mode handles considerably more [2].

This is a quiet but important admission from the industry: diarization at scale — the classic 8-10 person meeting room — is still the hard problem nobody's fully solved. Smallest.ai is explicitly targeting the B2B infrastructure layer underneath tools like Fireflies, Otter, Fathom, and Granola, positioning Pulse as the engine rather than the app [3].

X reactions leaned enthusiastic about accuracy gains in larger meetings, though it's worth treating "high accuracy for 8-10 speakers" claims with some skepticism until independent benchmarks land. Still, if diarization infrastructure keeps improving at this pace, the gap between "who said what" accuracy in a 2-person call versus a 10-person all-hands should keep narrowing fast.

IBM Pushes Agentic Knowledge Graphs as the New AI Memory Layer

IBM, together with edX and Pluralsight, released a roughly one-hour course on building agentic knowledge graphs — covering entities, relationships, ontologies, and how agents extract and integrate structured and unstructured data into memory for multi-agent systems [1]. It sits alongside IBM's broader agentic AI curriculum on LangGraph, ReAct, and orchestration, with tooling references to Neo4j and LangChain for explainable, multi-hop reasoning [2].

The subtext is bigger than a training course: knowledge graphs are being repositioned not as a nice visualization layer but as the actual memory substrate for AI agents that need to reason over time, context, and relationships rather than just retrieve similar-sounding text. That's a meaningfully different architecture than a vector database stuffed with embeddings.

X discussion zeroed in on the PKM angle — people drawing lines from these enterprise frameworks to personal knowledge tools like Obsidian, suggesting the same principles (entities, relationships, explainable retrieval) apply whether you're building a company brain or a personal one [3].

HydraDB Launches Time-Aware Context Graphs for Enterprise AI Memory

HydraDB released what it's calling a "missing context layer" for AI: temporal knowledge graphs that treat time as a first-class citizen, with validity windows, timestamps, and Git-style append-only versioning [1]. The system combines hybrid retrieval — vector search, graph traversal, and recency weighting — with automatic entity and relationship extraction at ingestion time, aimed squarely at enterprise teams trying to track how facts, people, and decisions evolve [2].

The benchmark numbers are notable: 90.79% accuracy on LongMemEval-S using Gemini 3.0 Pro, with sub-200ms retrieval times, claimed to outperform standard vector databases specifically on temporal reasoning tasks [3]. That's the crux of the pitch — plain vector search is good at "what's similar" but bad at "what changed and when," which is exactly the kind of question that matters most when you're trying to recall what was decided three meetings ago versus what's true today.

X posts framed HydraDB explicitly against vector DBs, pitching it as the right substrate for remembering people, projects, and shifting context over long AI-assisted work relationships — not just single-session recall.

What This Means For Your Meetings

Four stories, one theme: the industry is converging on the idea that a meeting transcript is worthless without structure, memory, and time. Google's offline push proves users want their conversations to stay private by default. Smallest.ai's diarization gains show the foundational "who said what" problem is still being actively solved, not finished. And both IBM's course and HydraDB's launch make the same argument from different angles — that knowledge graphs, not flat transcripts or embeddings, are what let AI actually reason over your history rather than just search it.

This is exactly the territory Proudfrog has staked out from day one. A transcription tool that only transcribes is table stakes in 2026; the real value is in what happens after the meeting ends — turning dialogue into a structured, speaker-attributed, time-aware knowledge graph that your AI assistant can traverse months later when you ask "what did we agree with that vendor back in March, and has anything changed since?" That's a temporal reasoning question, not a keyword search, and today's news confirms the whole industry is racing toward graph-based memory to answer it.

For professionals drowning in back-to-back calls, the practical implication is this: the tools that win long-term won't be the ones with the cleanest live transcript, but the ones that let you retrieve, connect, and trust what was said — privately, accurately, and with full context of how your projects evolved over time.

Key takeaway: Transcription is becoming commoditized — privacy, diarization accuracy, and time-aware knowledge graphs are where the real competitive battle for your meeting intelligence now lives.

Sources

1. https://www.theverge.com/tech/1007985/google-ai-notetaking-app-transcribe-offline
2. https://www.cnet.com/tech/services-and-software/google-ai-note-taker-edge-foresight/
3. https://www.notebookcheck.net/Google-AI-Edge-Foresight-Offline-AI-for-meeting-notes-on-Mac.1418337.0.html
4. https://docs.smallest.ai/models/speech-to-text/features/diarization
5. https://smallest.ai/speech-to-text
6. https://docs.smallest.ai/waves/model-cards/speech-to-text/pulse
7. https://www.pluralsight.com/courses/agentic-knowledge-graphs
8. https://www.edx.org/learn/computer-science/agentic-ai-with-langchain-and-langgraph
9. https://developer.ibm.com/technologies/agentic-ai/learningpaths/
10. https://hydradb.com/blog/temporal-knowledge-graphs-tracking-how-ai-context-evolves-over-time
11. https://docs.hydradb.com/essentials/v2/context-graphs
12. https://hydradb.com/blog/knowledge-graph-memory-systems-ai-agents