Google and Arm have published a more concrete recipe for running audio generation on-device with LiteRT, using Stability AI's Stable Audio Open Small as the reference model.

What shipped

In a May 14 engineering post, Google said developers can convert the PyTorch model with LiteRT Torch, inspect and tune quantization choices with Model Explorer and AI Edge Quantizer, then deploy the resulting .tflite models through LiteRT's runtime on Arm CPUs. Arm's companion sample repository publishes the full audiogen workflow, including conversion scripts and app build instructions for Android, Linux, and macOS.

Google's write-up focuses on the model's diffusion-transformer submodule, where the team says dynamic INT8 quantization improved performance and cut memory use without materially hurting output quality. In the benchmarks published by Google and Arm, generating an 11-second audio clip dropped from about 10 seconds to 4.3 seconds on an M4 MacBook, and from about 14 seconds to 6.6 seconds on an Arm SME2-based Android device using one thread.

Why it matters

This is not a new consumer product launch. It is developer plumbing for a hard edge-AI problem: getting multimodal generation to run locally on ordinary CPU-centric devices. The conservative takeaway is that Google is trying to make LiteRT more than an inference runtime by pairing conversion, visualization, quantization, and deployment into a repeatable path for small generative models.