FUTO's ASR4ALL Targets CPUs, Raspberry Pis, and the Edge

FUTO's ASR4ALL Targets CPUs, Raspberry Pis, and the Edge

A day before Tether's release, FUTO announced ASR4ALL — a family of streaming speech models in small (28M), medium (56M), and large (97M) parameter sizes, engineered specifically for low-power CPU inference [4]. They ship in GGUF, ONNX, and ExecuTorch formats, run on hardware as modest as a Raspberry Pi or an Intel integrated GPU, and the fastest configuration hits 73,000x real-time factor on GPU.

The pitch is explicitly about the Pareto frontier: near state-of-the-art accuracy without state-of-the-art hardware requirements. Combined with Tether's release, it's a clear signal that the ASR research community is racing toward "transcription everywhere" — cheap, fast, offline, and no longer gated by expensive inference infrastructure.

LeCun Takes Another Swing at Proprietary AI Labs

Yann LeCun used X on October 16 to call proprietary AI model companies "the IBMs of yesteryear," pointedly noting that the new IBM — in his telling, Meta — is the one backing open source [5]. It's a continuation of his long-running argument, newly sharpened amid his broader commentary on AI's shift toward physical-world applications and where strategic advantage will actually accrue [6].

Reactions were split, predictably, but the sovereignty argument is gaining traction in enterprise and government circles: organizations wary of concentrating control (and data) in a handful of closed labs are increasingly citing open-weight models as the pragmatic choice, not just the ideological one.

Agentese Pushes AI Agents Into Live Meetings, Not Just Summaries

Agentese launched a suite of task-completing AI agents, headlined by a Meeting Copilot that joins Google Meet calls live, answers questions in real time using meeting context and web access, and operates across multiple languages [7]. Other agents in the lineup handle phone calls, job applications, approvals, and code audits — the common thread being action during the event, not retrieval after it [8].

The buzz on X centers on a real shift in expectation: why wait for a post-call summary when an agent can answer a factual question live, mid-discussion? It's a preview of where meeting intelligence tools are headed — from passive recorders to active participants.

EU AI Act Clarifies Its Light Touch on Open-Source Models

On the regulatory side, clarifying analysis this week confirmed that the EU AI Act, phasing in through 2027, largely exempts open-source AI models from provider obligations — unless they're classified high-risk under Annex III, carry transparency obligations, or qualify as systemic-risk general-purpose AI [9]. Open-source GPAI providers still owe training data summaries and copyright compliance, but the compliance bar is meaningfully lower than for proprietary systems [10].

For Nordic and broader European enterprises, this matters concretely: fine-tuning an open model for a regulated use case (health, finance, legal) triggers full obligations regardless of license, so "open-source" isn't a blanket shield — but for general deployment, it's a lighter regulatory path than many assumed [11].

What This Means For Your Meetings

Today's news cluster points in one direction: transcription is becoming commoditized, fast, and local. When a 254M-parameter model can transcribe ten minutes of audio in under two seconds on a phone, with accuracy beating models many times its size, the bottleneck in meeting intelligence stops being "can we transcribe accurately" and starts being "what do we do with everything we've transcribed." That's where knowledge graphs, speaker identification, and cross-meeting retrieval — not raw transcription — become the actual differentiator.

The Agentese news is the other half of this story. Live, in-meeting AI participation (answering questions mid-call, pulling in context from the web or from history) is the next frontier beyond passive note-taking. Combine that with on-device models that never send audio to the cloud, and you get a credible path toward meeting tools that are both more capable and more private — exactly the tension Nordic and European enterprises have been navigating under GDPR and the AI Act's open-source exemptions. Regulatory clarity on open models lowers the barrier for building compliant, sovereign meeting intelligence stacks rather than depending on a handful of proprietary API providers, echoing LeCun's broader point about concentration of control.

For teams building or buying meeting intelligence tools, the practical implication is this: raw transcription accuracy is rapidly becoming table stakes, available even offline and on cheap hardware. The real value — and the real product differentiation — lives in what happens after the words are captured: how they're linked across people, projects, and time, and how reliably you can ask your own meeting history a question and get a trustworthy answer back.

Key takeaway: As transcription becomes free, fast, and local, the competitive edge in meeting tools shifts entirely to retrieval, memory, and live participation — not who can produce the most accurate transcript.

Sources

1. https://qvac.tether.io/blog/on-device-speech-recognition-transcribepsy
2. https://blockchain.news/news/tether-releases-on-device-speech-model-for-smart-glasses-and-robots
3. https://github.com/tether-ai-research/qvac-transcribepsy
4. https://futo.org/blog/asr4all/
5. https://x.com/ylecun/status/2111074492592091264
6. https://www.koreaherald.com/article/10899735
7. https://x.com/AgenteseAI/status/2037418270274498880
8. https://x.com/Agentese_AI/status/2043501985622860072
9. https://presencis.com/questions/ai-act-open-source/
10. https://wavenetic.com/blog/open-source-ai-models-for-regulated-industries-europe
11. https://naaia.ai/en/ai-act-modeles-ia-open-source-usage-general/