The niagara-19m-batch.en model is a State Space Model (SSM) with attention that performs automatic speech recognition (ASR), trained and released by Applied Brain Research (ABR). This model contains ~19m parameters, transcribes speech in English, and was trained on ~50k hours of speech data (competitors use about 200k hours). This is a batch model variant designed for direct comparison with other models on the Open ASR Leaderboard. SSMs are an ideal solution for streaming contexts, but to provide a more direct comparison with other models in the leaderboard this model is not streaming. A live streaming variant of this english ASR model with sub-120ms latency and features such as punctuation, capitalization and custom vocabulary along with other similar models for other languages is available under commercial license from ABR.
Why “Niagara”?
ABR names its model families after rivers. State space models process sequential data as a continuous flow, much like a river — always moving forward, maintaining state efficiently over time. Niagara is ABR’s ASR model family, named for the river in our home province of Ontario, Canada.
audio_list = [sample["audio"]["array"] for sample in samples]
batch_features = feature_extractor(audio_list)
batch_outputs = model(batch_features["input_features"], mask=batch_features["mask"])
transcriptions = tokenizer.decode_from_logits(
batch_outputs["logits"], mask=batch_outputs["mask"]
)
for t in transcriptions:
print(t)
Input
This model accepts 16000 Hz Mono-channel Audio (wav files) as input.
Output
This model provides transcribed speech as a string for a given audio sample.
Model Details
The SSM ASR model is trained for English speech recognition and transcribes audio into text. ABR developed the model to demonstrate small, efficient, real-time, accurate speech recognition can be performed with SSMs and run on low cost third party hardware. The model uses 19m parameters. The version posted here is a non-causal model (like most on the leaderboard), to give fair performance comparisons. It is also available as a cascaded model, meaning it produces extremely low latency (<120ms from first audio to token) causal outputs as well as 1s latency non-causal outputs. These two streams can be merged to have a quick response that updates after 1s with the final result.
Release Date: April 15, 2026
Model Type
Automatic speech recognition model transcribing speech audio to text in English.
Model Use
The intended use of the model is for evaluation by AI developers who want extremely small but performant ASR. We recognize that it is not possible to enforce our intended use guidelines. The models should not be used to transcribe individuals without their explicit consent, or be used to infer any particular human features as only text output is generated by the model. Other capabilities have not been evaluated. We recommend against using the model in high-risk settings (such as making important decisions) where errors in the model output can result in significant consequences for users. We strongly recommend that users perform extensive evaluations for their use cases.
Training
The model was trained on datasets partially listed below. It uses MFCC preprocessing on the input and is trained with CTC loss. It uses greedy CTC decoding and sentencepiece tokenization.
Datasets
The training datasets contain ~50k hours of English speech, including:
LibriSpeech (clean)
VoxPopuli
GigaSpeech
Common Voice
TED-LIUM
Europarl
Earnings-22
AMI-IHM
SPGISpeech
Performance
The posted model transcribes audio to lower case english with no punctuation. Performance is reported in terms of Word Error Rate (WER%) for the non-causal model.
Benchmark
niagara-19m
niagara-38m
AMI
18.28%
15.46%
Earnings22
13.18%
11.83%
GigaSpeech
14.21%
11.41%
VoxPopuli
9.82%
8.67%
AMI (Cleaned)
16.71%
13.76%
Earnings22 (Cleaned, AA-chunked)
16.33%
12.94%
GigaSpeech (Cleaned)
14.07%
11.31%
LibriSpeech test.clean
4.45%
3.75%
LibriSpeech test.other
10.95%
9.22%
SPGISpeech
3.86%
3.08%
VoxPopuli (AA-Cleaned)
5.93%
5.16%
Average WER
11.62%
9.69%
RTFx
3,710
4,015
Word Error Rate (WER), lower is better. RTFx is inverse real-time factor, higher is
better. All figures from the Open ASR Leaderboard,
measured on NVIDIA A100-SXM4-80GB at batch 64 where memory allows. Average is the
mean of the eleven columns above and reproduces the leaderboard's Average WER.
Figure 1: ASR Leaderboard Data accuracy vs parameter count. Transformer-based models marked in grey. SSMs improve the pareto front of transformers.
Models with fewer than 100M parameters, sorted by accuracy.
Model Name
Parameters (M)
Accuracy (%)
abr-ai/niagara-38m-batch.en
38
91.09
moonshine-base
61
90.00
whisper-base
74
89.70
abr-ai/niagara-19m-batch.en
19
89.53
moonshine-streaming-tiny
34
88.00
moonshine-tiny
27
87.30
whisper-tiny
39
87.30
Limitations
Since this model was trained on publicly available speech datasets, the performance of this model might degrade for speech which includes technical terms, significant noise, or vernacular that the model has not been trained on. The model might also perform worse for accented speech. The model might generate text that was not actually spoken in the input audio.
Broader Implications
We intend for the SSM ASR model to be used for beneficial purposes, including low cost transcription on low cost hardware, providing accessibility, and improving real-time voice user interfaces. As with all AI technology, there are also reasons to be concerned about dual-use. For instance, lowering the cost may allow broader deployment of undesired surveillance technology or the inexpensive scaling of existing technology. Related safety concerns come from the model being used to identify individuals or being deployable in very small footprint hardware.
This open-source batch model demonstrates ABR’s SSM architecture for ASR. ABR also offers commercially licensed models with additional capabilities: real-time streaming inference (sub-120ms first-token latency from first audio), speech recognition for other languages, punctuation and capitalization, custom vocabulary, hardware-optimized deployment for many leading edge processors, Text-to-Speech (TTS) models, and enterprise support with SLAs. For more information, visit appliedbrainresearch.com or contact info@appliedbrainresearch.com.
About Applied Brain Research
Applied Brain Research (ABR) is based in Waterloo, Ontario, Canada. ABR develops efficient speech and language AI using its proprietary and patented state space model architectures. More at appliedbrainresearch.com.
