Stability AI and Arm bring offline generative audio to smartphones

Need a quick whoosh sound, rain ambience, or a cinematic boom for your mobile video project? Usually, you either scour royalty-free libraries or ping an expensive cloud server. If you are stuck in airplane mode, you are completely out of luck.

Stability AI wants to change that. On Monday at MWC Barcelona, the company announced a partnership with Arm to run its text-to-audio model, Stable Audio Open, entirely on-device and completely offline on standard smartphone hardware.

From four minutes down to single-digit seconds

Shrinking generative audio to run locally is notoriously tough on silicon. Initially, generating a short audio clip using Stable Audio Open on an Arm CPU took an agonizing 240 seconds—essentially useless for any real-world creator.

To bridge the gap, Stability AI distilled the model, while Arm brought in its optimized software stack. By leveraging int8 matrix multiplication kernels from Arm KleidiAI inside ExecuTorch via XNNPack, the team pulled off a massive speedup.

Running on Armv9 CPUs, generating an 11-second audio clip now takes under 8 seconds. That is a 30x performance boost, achieved purely on the CPU without requiring a specialized, proprietary NPU.

Stability AI says audio is just the opening act. The company plans to bring its full suite of generative media models—including image, video, and 3D generation—directly to edge devices.

Why it matters

For mobile creators, indie game developers, and video editors, having zero-latency, offline sound effect generation right on your phone is a huge win. You get custom foley and ambient tracks without burning mobile data, waiting on server queues, or paying recurring cloud API fees.

For the broader AI ecosystem, this marks a shift in how edge AI workloads get distributed. While most mobile AI marketing has focused on on-device LLMs or basic photo cleanup, generative audio was largely written off as too compute-heavy for local chips. Proving that an Arm CPU can generate high-fidelity audio in seconds shows that clever kernel-level engineering can rival dedicated silicon.

We think partnering with Arm is Stability AI's smartest distribution move in a while. Arm tech powers virtually 99% of global smartphones. By tapping into Arm's KleidiAI libraries, Stability AI avoids the headache of negotiating proprietary NPU toolchains across different phone brands.

Still, there are practical hurdles ahead. While 8 seconds for an 11-second sound bite is usable, it is far from real-time streaming. Running continuous heavy int8 workloads on a mobile CPU will inevitably chew through battery life and kick up thermal throttling on thinner devices. It is a massive technical leap, but creators will likely treat it as an occasional utility rather than a continuous workflow engine until thermals catch up.