In my previous post, I introduced State Space Models (SSMs) and explained why they matter for edge AI. Today, we’re putting that theory into practice with Niagara-38m — a new ASR model that sets a new best-in-class benchmark for efficiency, accuracy, and real-time performance. Hugging Face has now published this new standard by posting the Niagara-38m results on their ASR Leaderboard.
At 38 million parameters, ABR’s new English ASR SSM achieves accuracy within 0.3 percentage points of OpenAI’s whisper-small — a model six times bigger. ABR also offers a commercial version that provides true streaming audio with ~150ms latency (first audio in to first token out). And both run on hardware that whisper-small cannot reach.
This is not just a benchmark result. It means real-time voice interfaces can now run directly on devices like wearables, AR glasses, and robots without relying on the cloud. This is what we said SSMs would make possible. Here’s the evidence.
Setting a new bar on the leaderboard
The figure below tells the story. ABR’s Niagara SSM models sit above the Transformer Pareto frontier — the boundary that marks the best accuracy achievable per parameter by transformer-based architectures. Every point in the grey zone is a transformer model. The ABR models set a new frontier boundary for efficient accuracy.
Figure 1: ASR Model Leaderboard — Accuracy vs. Parameter Count. Data as of June 11, 2026. Niagara-10m not yet submitted. Full leaderboard: Open ASR Leaderboard
The most useful comparison is to the industry standard whisper family from OpenAI. Niagara-38m achieves an astounding accuracy of 90.64% that:
- Is nearly four percentage points higher than that of whisper-tiny, which has a million parameters more.
- Beats whisper-base, with almost twice the number of parameters but still doesn’t reach 90% accuracy.
- Comes within 0.3 percentage points of whisper-small which has 244 million parameters. That’s over six times the size!
Niagara-38m also clocks in with a real-time factor (RTF) of 2623, which is 7.5x faster than whisper-base at 348. Better accuracy and faster inference together — each a reflection of the same architectural efficiency — make this model deployable at the edge at performance levels transformers can’t reach.
Following suit, ABR’s Niagara-19m model is also on the leaderboard at 88.99%, surpassing whisper-base while being nearly a quarter of its size. We will also shortly release Niagara-10m on Hugging Face, which is also shown in the above figure. Together, the Niagara models set a new performance standard beyond the reach of transformers.
Streaming models for the edge
The leaderboard uses the ‘batch processing’ version of the Niagara model — the same execution mode as competing models. It’s the right basis for comparison, but it isn’t well-suited for edge devices that need to operate in real time. As I pointed out in my last blog post, SSMs are natively streaming in nature and don’t suffer from a need to batch process to achieve accuracy.
ABR’s Niagara-live line of models, which are available under commercial license, are natively real time. These models produce two output streams at once: a ‘causal’ stream that produces output immediately with world-leading average first-token latency of ~150ms, and a ‘non-causal’ stream that arrives about one second later with slightly higher accuracy. The outputs can be merged or used independently, whatever best fits the needs of the application. When merged, the user sees an immediate result that quietly self-corrects at the one-second mark. Fast first, accurate always. There is no transformer-based equivalent at this scale.
This matters anywhere latency is felt: voice-first interfaces, live captioning, real-time dictation. For any product where the interaction needs to feel responsive, this “first audio to first token” latency is the metric that counts and only ABR’s SSM technology is architected with it in mind.
In practice, this means users see results instantly, with accuracy improving in the background — creating an experience that feels both fast and reliable.
Delivering the complete product
The open model is the starting point — designed to show what the SSM architecture can do at 38 million parameters. But Niagara-live delivers far more than just real-time accuracy in a compact footprint.
We have packaged the model with a full set of features required by the market:
- Capitalization and punctuation — because raw lowercase output isn’t usable in documents, captions, or some interfaces; the commercial version produces text you can put in front of users directly.
- Custom vocabulary support — so the model handles domain-specific terminology out of the box without retraining.
- Additional languages beyond English — because edge devices need to operate in a multilingual world.
ABR is also releasing a native SDK in the coming weeks with bindings for C, Python, and Java. It’s designed for drop-in integration: no custom ML pipeline, no internet connection, no model-serving infrastructure required. The SDK is what lets you quickly turn this from a model you evaluate into a product you ship.
What this means
Accurate, efficient, deployable speech recognition no longer requires a large model, a fast connection, or a cloud contract. At 38 million parameters, Niagara-38m fits where the next generation of voice products actually lives — on the device, at the edge, in the hands of the user.
And as explained in our blog post on cloud voice AI costs, integrating this model and its text-to-speech counterpart with an LLM can reduce the cost of a streaming voice integration by 10-40x. Perfect anywhere you want to give users a natural voice interface.
Try Niagara-38m today on Hugging Face. For more info on Niagara-live — streaming, capitalization and punctuation, custom vocabulary, and the SDK — contact us directly.