Four open Niagara models, from 9M to 84M parameters, are the most accurate at their size among models under 100M parameters
Update, October 28, 2026: Models added to the leaderboard on October 26 now share the accuracy-to-size frontier below 100M parameters. The Niagara models still hold their positions on the frontier which they had defined since Sept 25.
October 26, 2026. Applied Brain Research (ABR) today announced that its four open Niagara batch speech recognition models, at 9M, 19M, 38M and 84M parameters, are listed on the Hugging Face Open ASR Leaderboard and define the Pareto frontier of word error rate against parameter count for models under 100M parameters. That is, for models in that range, the most accurate model at each size on the leaderboard is a Niagara model. These models are built on the state space model architecture invented by ABR, which delivers higher accuracy with fewer parameters than other models under 100M parameters on the leaderboard.
The Open ASR Leaderboard, maintained by Hugging Face, evaluates speech recognition models on a common set of English test sets, including public datasets and private data held by the maintainers, with results computed by the maintainers under the same conditions.
Parameter count largely determines the memory, compute, and power a speech model needs on a device. Wearables, hearables, AR glasses, appliances, and other edge products often have tens of megabytes of memory or less. Niagara models are built with the needs of these devices as the central focus, delivering high accuracy with low demands on compute and memory.
“Accuracy at a given model size is the measure that matters for on-device speech,”
said ABR CTO Dr. Chris Eliasmith.
“The models we compare against were built for edge applications, but they use transformer architectures that were designed for the abundant compute and memory of data centers. These results show that a state space model architecture, designed from the start for continuous signals, delivers more accuracy per parameter.”
The batch models on the leaderboard are available for evaluation on Hugging Face. ABR’s commercial streaming Niagara models use the same architecture and are delivered through the ABR SDK, together with ABR’s Nith text-to-speech models. All ABR streaming models run faster than real time and support continuous streaming on supported platforms.
Availability
Open ASR Leaderboard: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard
niagara-9m-batch.en: https://huggingface.co/abr-ai/niagara-9m-batch.en
niagara-19m-batch.en: https://huggingface.co/abr-ai/niagara-19m-batch.en
niagara-38m-batch.en: https://huggingface.co/abr-ai/niagara-38m-batch.en
niagara-84m-batch.en: https://huggingface.co/abr-ai/niagara-84m-batch.en
Documentation: https://docs.appliedbrainresearch.com
Developer portal: https://dev.appliedbrainresearch.com
About Applied Brain Research
Applied Brain Research develops and licenses AI software that uses its proprietary state space models specifically designed for embedded edge devices, enabling speech recognition, text-to-speech, translation, and natural language processing to run on-device within tight power and memory budgets. For more information, visit www.appliedbrainresearch.com.
Media contact
Kevin Conley, CEO, [email protected]