Tiny Audio AI Models Learn to Whisper Without Ever Calling Home

Samsone: Sub-400M Audio Models Aimed at Phones

A research team has posted Samsone, a family of small audio language models built specifically for on-device speech tasks. The lineup spans 99M, 134M and 356M parameters, with the 134M version claimed as a new state of the art for its size class across multiple benchmarks — hundreds of times smaller than the billion-parameter audio models usually cited in these comparisons.

The team released training code, model weights, mobile-optimized checkpoints and an open-source Android app so anyone can test real-time on-device listening. The benchmark claims are the paper's own; nobody outside the project has reproduced them yet, and 'competitive with much larger models' is a common self-reported line in this space that deserves independent replication.

Still, an app-ready demo plus open weights at under 400M parameters is a real data point for anyone trying to run speech recognition on a phone without a network call. Coverage via The Neural Feed.

Applied Brain Research Ships a Streaming Voice SDK That Stays on the Device

Applied Brain Research announced general availability of the ABR SDK on September 21, 2026, bundling its Niagara streaming ASR and Nith streaming TTS models behind one API so both speech-in and speech-out can run on embedded processor hardware with no network dependency.

The company reports first text from Niagara in as little as 115 milliseconds and first audio from Nith in as little as 147 milliseconds, measured on embedded application-class CPUs — vendor benchmarks, not independently verified. Five languages ship at launch (English, Spanish, Mandarin, Japanese, Korean), with a Python library over a stable C ABI now and C/Java bindings promised soon.

Each model comes as a self-contained directory of weights plus config, so swapping language or upgrading weights means changing a file path rather than recompiling. That packaging choice matters more for shipping products than the latency numbers do; details via Physical AI News.

Cactus Compute's Needle Grows a Third Version, Still Fits Under 30MB

Cactus Compute has followed up its earlier sub-30MB tool-calling model with Needle 3, a new checkpoint shipping as a single file between 8MB and 29MB depending on configuration, aimed at phones, wearables and microcontrollers.

The model narrows its scope to tool calling, structured extraction and text embedding rather than general chat, and the company claims it beats models ten times its size on mobile tool-calling benchmarks — a self-reported figure with no independent check available yet. It installs via pip install cactus-needle, with a 20-layer inference file, a safetensors checkpoint for fine-tuning, and platform engines each under 1MB on the project's repo.

The repo logs over 46,000 downloads in the past month, though no third-party inference provider has deployed it yet. It is a narrow bet — trade generality for a model small enough to sit on a microcontroller — but a plausible one for agents that mostly need to call functions, not converse.

Nordic's Smallest BLE Chip Sleeps at Under 50 Nanoamps

Nordic Semiconductor has expanded its nRF54L series with the nRF54LC10A, a Bluetooth LE and 802.15.4 SoC trimmed down in memory and peripherals from the higher-end nRF54L parts, reported September 22, 2026 by CNX Software.

The headline spec is a sub-50 nanoamp hibernation mode, aimed squarely at coin-cell wearables and sensor nodes that spend most of their life asleep. Exact flash and RAM sizes, and pricing, weren't fully detailed in the initial listing, so this is a spec worth watching until Nordic publishes the full product brief.

It's not an AI chip on its own, but ultra-low hibernation current is exactly the constraint that shapes whether a wearable can run periodic on-device inference at all without draining a battery in days.

Zephyr 26.04 Lands With a 15-Year Support Window

Canonical announced Zephyr 26.04 LTS on September 21, 2026, an enterprise distribution of the RTOS with up to 15 years of security maintenance and Golioth's OTA update mechanism built in by default.

Long support windows matter more for embedded than for phones or PCs: a device shipped once and left in the field for a decade needs security patches without a full re-certification cycle. Bundling OTA into the LTS baseline lowers the bar for smaller teams to ship update infrastructure they'd otherwise have to build themselves.

None of this runs a model directly, but it's the kind of firmware plumbing that determines whether the ESP32-class and Cortex-M boards carrying tiny LLMs and voice pipelines stay patchable years after they leave the workbench.

The voice stack is quietly finishing its move on-device — small audio models, streaming SDKs, and narrow tool-calling checkpoints are converging on the same target: a phone or wearable that never has to phone home to understand or answer speech. The chips and firmware underneath are getting more patient batteries and longer support windows to match.