Models

On-device speech models, designed for integration

What integration teams need to know about our models: architecture, target platforms, and how the SDK and per-language adapters fit together.

Architecture

Our models are built on Whisper Tiny, a compact speech recognition model published by OpenAI. We adapt it with speech data recorded in our own studios. Adaptation changes the weights, not the architecture: each model keeps its size.

Everything runs on the device. No audio leaves it, and no network connection is needed.

Base modelWhisper Tiny (OpenAI), encoder-decoder transformer
TaskSpeech recognition
Parameters37.2 million
SizeAbout 75 MB (FP16)
ExecutionOn the device, offline

Target platforms

Our models run on Android smartphones (ARM64, Android 8.0 or later), offline and with no network permission. They are built for these device families:

  • Smartphones
  • Tablets
  • Computers
  • TVs
  • Cars
  • Home appliances
  • Smart speakers

Tell us your platform and chipset: we adapt the models to your hardware.

SDK and per-language adapters

Your productYour application
On the device, offline
Labari SDK · on-device runtime
Base model
fr-BJ adapter fon-BJ adapter wo-SN adapter + more languages
Adding a language means shipping a new adapter; the integration stays the same.
  1. One runtime per platform. The SDK runs the base model on the device and exposes a single call: audio in, text out.
  2. One adapter per language. Each language or local variety is delivered as an adapter, a set of weights trained on our data and loaded on top of the base model.
  3. New languages without a new integration. Adding a language to a device range means shipping a new adapter. The integration stays the same.