Waterloo, ON – September 21, 2026 -- Applied Brain Research (ABR) has released the ABR SDK, a production toolkit combining its Niagara streaming speech recognition and Nith streaming speech synthesis models into a single API that runs entirely on edge processor hardware, eliminating network dependency for voice interfaces.
Niagara ASR generates first text output in as little as 115 milliseconds
The Niagara automatic speech recognition models produce initial text from audio input in 115 milliseconds, measured on embedded application-class CPUs. The Nith text-to-speech models generate first audio output from initial text in 147 milliseconds, with all models operating faster than real time across every supported platform.
Voice data stays on-device, removing cloud dependency for latency-critical processing
Because speech input and output processing runs locally, voice data does not leave the device and the interface continues functioning during unreliable connectivity. "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.
SDK covers five languages with a single Python library over a stable C ABI
The current release supports English, Spanish, Mandarin, Japanese and Korean, with C and Java bindings scheduled in the coming weeks. Each model ships as a self-contained package bundling the compiled library, model weights and configuration, so switching languages or upgrading weights requires only a path change rather than a cross-compilation step.
Voice cloning and custom vocabulary target brand- and domain-specific deployments
An add-on voice cloning feature generates a new synthesis voice from a short reference recording using consent-based training. A custom vocabulary capability allows domain terms, proper nouns and product names to be recognized and pronounced correctly by both ASR and TTS models without retraining.
Platform support spans Linux x86-64, Linux ARM64 and Android ARM64
Acceleration on integrated NPUs and DSPs is available on certain partner platforms, with additional silicon targets and RTOS support for Cortex-M class microcontrollers with NPUs targeted before year end.
Niagara batch model leads Hugging Face Open ASR Leaderboard under 100 million parameters
ABR's open non-streaming niagara-38m-batch.en model holds the top accuracy ranking under 100 million parameters on the Hugging Face Open ASR Leaderboard, at a mean word error rate of 9.69 percent as of the release date. The smaller niagara-19m-batch.en model outperforms models several times its size on the same leaderboard. The streaming SDK models share the same state space model architecture as these batch models and train on a superset of the same data.
Self-managed registration provides free evaluation access to the full SDK and available models through ABR's developer portal, with pilot and production deployment available under commercial license.