Applied Brain Research Releases the ABR SDK for On-Device Voice Interfaces to Edge Applications

Applied Brain Research (ABR) has released the ABR software development kit and the Niagara speech recognition and Nith text-to-speech model families, a production toolkit for building real-time voice interfaces that run entirely on edge processor hardware.

The SDK provides streaming automatic speech recognition and streaming text-to-speech through a single API, so developers can enable responsive voice interfaces on devices. Because the latency-critical speech input and output processing runs locally, there is no speech AI dependency on a network, voice data does not leave the device, and the interface continues to operate when connectivity is unreliable.

"A voice interface is only usable if it answers immediately, which makes real-time performance the binding constraint for edge applications," said Kevin Conley, CEO of Applied Brain Research, in a statement. "ABR's models and SDK are designed to deliver leading accuracy with the lowest latency on constrained edge hardware. Customers and hardware partners evaluating the SDK are consistently surprised by the response times and accuracy they measure."

The current release provides a Python library over a stable C ABI. A single SDK covers ABR's Niagara speech recognition and Nith speech synthesis families running simultaneously, starting with English, Spanish, Mandarin Chinese, Japanese, and Korean.

Each model is distributed as a self-contained package holding the compiled library, the model weights, and the configuration in one directory. Changing language or upgrading to newly trained weights requires only a change of the path.

Two add-on capabilities are available to address brand- and domain-specific applications. For TTS, voice cloning generates a new synthesis voice from a short reference recording using consent-based training. For ASR and TTS, custom vocabulary allows a product's domain terms, proper nouns and product names to be recognized and pronounced correctly by both the recognition and synthesis models, without retraining.

ABR's Niagara streaming ASR models produce first text from the initial audio in as little as 115 milliseconds. The Nith streaming TTS models produce first audio from initial text in as little as 147 milliseconds, measured on embedded application-class CPUs.

Linux x86-64, Linux ARM64 and Android ARM64 are currently supported. Acceleration on integrated NPUs and DSPs is also available on certain partner platforms. Additional silicon targets and support for RTOS running on Cortex-M class microcontrollers with NPUs are targeted for release before year end.