Both fidelity yardsticks are reported so the codec is judged fairly: faithfulness is PRD vs the recorded (still-noisy) input, while truth fidelity is PRD vs clean ground truth. Lower is better.

Metric	Value
Truth PRD vs clean (%)	7.28
Truth PRD at native noise (%)	38.22
Faithful PRD vs input (%)	9.72
PRD degradation slope (PRD%/dB)	3.86
PRD at 0 dB SNR (%)	49.37
PRD at -6 dB SNR (%)	77.25
Pure-noise imprint autocorr	0.2637

Time Domain

PRD here is faithfulness (vs the recorded input); see Fidelity & Robustness above for the clean-truth and noise-regime view.

Metric	Mean	Median	P90
PRD vs input — faithfulness (%)	9.7245	8.2869	16.0694
RMSE	0.0933	0.0793	0.1507
Cosine Similarity	0.9937	0.9966	0.9987

Spectral

Band Total Relative Error (median): 0.1178

Bitrate

Codec CR (uniform): 16.0x
Codec CR (learned prior): 26.88x

Encoder Precision Parity

Difference from FP32 reconstruction on a disjoint real-data holdout; lower is better.

Encoder	P90 PRD	Worst PRD	Status
FP16	1.50%	4.05%	recommended
INT16X8	5.07%	16.38%	recommended
INT8	9.13%	17.10%	recommended

Usage

Python (compressionkit runtime)

python

from compressionkit.runtime import RVQCodec

codec = RVQCodec.from_pretrained("Ambiq/compressionkit-ecg-16x-v1.0")

# Encode: float32 signal → RVQ indices
indices = codec.encode(signal)

# Decode: RVQ indices → reconstructed signal
recon = codec.decode(indices)

Local deployment directory

codec = RVQCodec("path/to/deploy/")

Files

File	Description
`encoder_int8.tflite`	INT8 quantized encoder (on-device)
`encoder_float32.tflite`	Float32 encoder for browser/server runtimes
`encoder_fp16.tflite`	FP16 encoder variant for supported edge runtimes
`encoder_int16x8.tflite`	INT16x8 encoder variant for supported edge runtimes
`encoder.h`	C header for encoder
`encoder.keras`	Float32 Python reference encoder (training/inspection use)
`decoder_float32.tflite`	Float32 decoder (server-side evaluation)
`decoder_int8.tflite`	INT8 decoder (optional, on-device)
`decoder.keras`	Float32 Python reference decoder (training/inspection use)
`codebook.npz`	RVQ codebook tables
`codebook.h`	C header for codebook
`config.json`	Deployment manifest
`sample_stimulus.npz`	Synthetic test data
`demo_recordings.npz`	10 real, quality-gated browser-demo recordings
`demo_recordings_manifest.json`	Recording provenance, license, and quality metadata
`quality_scorecard.json`	Full evaluation metrics

Dataset & License

Training data: PTB-XL (CC BY 4.0). Sample data may include excerpts under the original license terms. Demo recordings: MIT-BIH Arrhythmia Database v1.0.0 (source); real, quality-gated excerpts are released under ODC-By-1.0. https://opendatacommons.org/licenses/by/1-0/

Model weights are released under the Ambiq Model Weights License — deployment is restricted to Ambiq silicon devices. See LICENSE-MODEL-WEIGHTS.md for full terms.

Citation

bibtex

@software{compressionkit,
  author = {Ambiq AI},
  title = {compressionKIT: Signal Compression for Edge AI},
  url = {https://github.com/AmbiqAI/compressionkit}
}