Custom AI models. Trained for the edge.
The on-device AI model training platform for enterprises. Train and fine-tune speech, language, and vision models for edge deployment. No ML team required.
On-device AI model training platform built for cross-platform enterprise deployments at scale
Training a machine learning model for on-device deployment is fundamentally different from training a model for the cloud. Processing data on the cloud requires optimization only for a certain platform or server. However, processing data on-device, i.e., various platforms from microcontrollers (MCUs) to web browsers, requires optimization for each platform. An on-device model has to fit within a strict RAM and CPU budget, and perform reliably across acoustic environments and various platform classes. General-purpose training frameworks don't account for any of this. They produce models that are accurate in the cloud and impractical at the edge. Due to platform variety, most on-device AI model training approaches cover only certain platforms, e.g., Apple optimizes for Apple products.
picoGym is built specifically for this problem. It handles the full machine learning model training pipeline — from transfer learning through hardware-aware architecture selection through evaluation — and outputs a model file that runs on any supported edge device, ready to integrate without conversion or post-processing.
What makes picoGym different from other on-device ML tools
Cloud AutoML and general-purpose training frameworks produce models optimized for server inference. Customizing models for edge deployment typically means hiring ML engineers, assembling datasets, configuring distributed training jobs, and debugging evaluation pipelines for weeks to achieve a production-ready model.
picoGym is vertically integrated from data pipeline through hardware-aware architecture through on-device runtime, handling all of this on a single platform. Because a model trained for the edge has to be designed for the edge from the start.
The only AI model training platform built for enterprise edge deployment at scale.
Why enterprises choose picoGYM On-device AI Model Training Platform
Train your first on-device AI model.
Your first custom model trains and deploys to any supported edge device without writing a single line of training code. No ML background required.
On-device AI model training is the process of building machine learning models that run inference directly on edge hardware — embedded devices, mobile or web apps, desktops, and local servers — without sending data to the cloud. picoGym automates this process: you define the target model type and deployment hardware, and the platform handles training, evaluation, and packaging.
Training time depends on model type. Models trained on Picovoice Console and in run-time, such as custom wake word models, speaker voice prints, voice commands, and speech-to-text models, are trained in under two seconds after the input is completed. Private training may take longer depending on the complexity of the model.
No. picoGym uses transfer learning from Picovoice's production model weights, so custom models reach production-grade accuracy without ML expertise, large datasets, or annotation pipelines. You define the target; the platform handles the rest.
Yes. Models trained via picoGym are designed for on-device inference with Picovoice Inference, without a runtime cloud dependency. Once deployed, they operate entirely locally — no API calls, no network requests, no data leaving the device. This makes them suitable for air-gapped infrastructure and applications with strict data residency requirements.
Cloud AutoML produces models that run in the cloud and introduce permanent API latency and data transmission requirements. General fine-tuning frameworks require ML expertise and produce models that need separate conversion and testing for each target hardware class. picoGym trains hardware-aware models from the start: architecture and quantization are chosen at training time for the target device, not added afterward.