Niagara Automatic Speech Recognition

Niagara Automatic Speech Recognition#

Niagara is a family of streaming ASR models optimized for local, real-time transcription. All processing runs on-device — from MCUs to GPUs — with no cloud dependency.

Built on a State Space Model (SSM) architecture with attention, Niagara achieves 150ms average latency from first audio to first token. Models range from 8M to 38M parameters, enabling deployment on resource-constrained edge hardware while maintaining accuracy competitive with much larger models. Supported languages include English, Spanish, and Mandarin Chinese.

Key Features#

  • Real-time streaming — Processing begins as soon as the first audio is received, minimizing latency

  • Private — All voice processing runs locally

  • Small footprint, high accuracy — 2–4× smaller than transformers at equivalent accuracy, verified independently on the HuggingFace Open ASR Leaderboard (see benchmark plot)

  • Cascaded output — Fast path (~150ms latency) for real-time display, refined path (~1s) for higher accuracy

  • Multi-language — English, Spanish, and Mandarin Chinese; additional languages available upon request

Premium Features

  • Custom vocabulary — Accurate identification of brand names, technical jargon, and domain- or application-specific words can be added without retraining the core ASR model

Performance Benchmarks#

Niagara is benchmarked on the HuggingFace Open ASR Leaderboard — an independently-run evaluation where all models are tested on shared datasets. Standardized datasets and public methodology enable direct comparison across models, where proprietary benchmarks which often lack sufficient detail for independent reproduction and can be misleading. Here, we report the performance of the batch variant of our Niagara models to match the leaderboard’s methodology.

ABR’s Niagara models (blue) achieve the highest accuracy among models under 50M parameters on the Open ASR Leaderboard. The 19M model outperforms Whisper-tiny at half the parameters; the 38M model matches the accuracy of transformer models 2–4× its size.

ASR Model Leaderboard — Accuracy vs. Parameter Count Transformers 87% 88% 89% 90% 91% 92% 93% 94% 95% 10M 100M 1B 10B Parameters (M) [log scale] Accuracy (%) niagara-19m-batch.en — 19M params, 89.53% accuracy niagara-38m-batch.en — 38M params, 91.09% accuracy whisper-tiny — 39M params, 87.30% accuracy whisper-base — 74M params, 89.70% accuracy whisper-small — 244M params, 91.40% accuracy whisper-medium — 769M params, 91.80% accuracy whisper-large-v2 — 1550M params, 92.20% accuracy Canary-180m-flash — 180M params, 92.80% accuracy parakeet-tdt-0.6b-v2 — 600M params, 93.50% accuracy canary-qwen-2.5b — 2500M params, 94.50% accuracy NVIDIA (100M) — 100M params, 91.00% accuracy NVIDIA (350M) — 350M params, 92.30% accuracy NVIDIA (600M) — 600M params, 92.50% accuracy NVIDIA (800M) — 800M params, 92.60% accuracy NVIDIA (1.5B) — 1500M params, 92.60% accuracy moonshine-tiny — 27M params, 87.30% accuracy moonshine-streaming-tiny — 34M params, 88.00% accuracy moonshine-base — 61M params, 90.00% accuracy Other — 80M params, 90.00% accuracy Other — 100M params, 92.20% accuracy Other — 110M params, 92.50% accuracy Other — 120M params, 91.40% accuracy Other — 150M params, 91.00% accuracy Other — 200M params, 93.30% accuracy Other — 250M params, 93.00% accuracy Other — 300M params, 92.80% accuracy Other — 320M params, 91.10% accuracy Other — 400M params, 93.30% accuracy Other — 500M params, 93.30% accuracy Other — 650M params, 92.30% accuracy Other — 800M params, 94.00% accuracy Other — 1000M params, 93.90% accuracy Other — 1100M params, 94.00% accuracy Other — 2000M params, 94.30% accuracy Other — 2800M params, 94.00% accuracy Other — 5000M params, 89.40% accuracy Other — 8000M params, 88.00% accuracy Series ABR NVIDIA OpenAI Whisper Other Data as of 11 March 2026
Word Error Rate (WER) measured across the Open ASR Leaderboard datasets. Models to the left are smaller, models higher up are more accurate.

Models#

Select a model based on your resource constraints. Larger models provide higher accuracy; smaller models run on more constrained hardware. All models accept 16kHz mono WAV audio and output punctuated, capitalized text.

38M 19M 9M 8M
Languages English ✓ ✓ ✓ ✓
Spanish ✓ — — ✓
Chinese (Mandarin) ✓ — — ✓
Sizes 32-bit 147MB 80MB 38MB 32MB
8-bit 39MB — 10MB 9MB

About the Name#

ABR names its model families after rivers. State space models process sequential data as a continuous flow — always moving forward, maintaining state efficiently over time. The Niagara model family is named for the river in ABR’s home province of Ontario, Canada, whose falls have the highest flow rate in North America.

Contents