opendsp

Numpy-only DSP primitives for biosignal analysis. A small, focused scipy.signal / PyWavelets subset implemented in pure NumPy, with optional backends (scipy, numba, torchaudio) for speed when available.

opendsp is the shared DSP layer underneath the open biosignal family (openvital, openecg, openeeg). The primitives here are modality-agnostic — bandpass filtering, peak-finding, wavelets, resampling — and don't belong to any single signal type.

Why

scipy.signal is excellent but heavy: scipy + its scipy.sparse / BLAS backing pulls 30+ MB and is awkward to install in stripped-down inference environments (Lambda, edge boxes, TFLite runners). opendsp matches scipy's coefficient output to within ~1e-12 on the filter orders biosignal pipelines actually use (2 and 4), so you can keep your filter design code identical while shrinking your deploy image.

Installation

pip install opendsp

# Optional speed backends (auto-selected on first lfilter call):
pip install opendsp[speed-scipy]   # ~10x speedup, C-backed
pip install opendsp[speed-numba]   # near-C after JIT warmup
pip install opendsp[speed-torch]   # GPU/MKL via torchaudio

Set OPENDSP_LFILTER_BACKEND (one of scipy, numba, torch, numpy) to force a specific backend — useful for benchmarking or reproducibility-sensitive contexts.

API

import opendsp

# Butterworth IIR design (scipy.signal.butter drop-in)
b, a = opendsp.butter(4, 0.1, btype="low")
b, a = opendsp.butter(4, [0.01, 0.4], btype="band")

# Direct filtering
y = opendsp.lfilter(b, a, x)
y = opendsp.filtfilt(b, a, x)   # zero-phase forward-backward

# Convenience bandpass (combines butter + filtfilt)
y = opendsp.band_pass(x, srate=128, fl=0.5, fh=30, order=4)

# Peak finding (scipy.signal.find_peaks subset)
peaks, props = opendsp.find_peaks(x, height=0.5, distance=10,
                                   prominence=0.1)

# Wavelets (PyWavelets minimal subset)
coeffs = opendsp.wavedec(x, "db2", level=4)
y = opendsp.waverec(coeffs, "db2")
cwt = opendsp.cwt(x, scales=[1, 2, 4, 8], wavelet="gaus1")

# Resampling
y = opendsp.resample(x, dest_len=1000)
y = opendsp.resample_hz(x, srate_from=500, srate_to=128)

# Misc primitives biosignal pipelines need
y = opendsp.interp_undefined(x)             # NaN-aware linear interp
y = opendsp.savitzky_golay(x, window_size=91, order=3)
y = opendsp.moving_average(x, N=10)
y = opendsp.rank_normalize(x)               # amplitude-invariant
y = opendsp.remove_baseline_wander(x, fs=500, cutoff_hz=0.5)

Algorithms

• Butterworth: analog prototype poles → bilinear-transform pre-warping → digital (z, p, k) → polynomial (b, a). Matches scipy exactly because both follow Oppenheim & Schafer §7.1.
• filtfilt: reflect-pad ↔ forward ↔ reverse ↔ forward ↔ reverse ↔ trim, with initial conditions set via the lfilter_zi steady-state method.
• find_peaks: strict-left/non-strict-right local maxima (scipy's convention), height filter, greedy-from-largest distance filter, Wim Spalt's two-sided prominence walk.
• wavedec/waverec: DWT with mode='symmetric' boundary extension (PyWavelets default). Daubechies-2 ships built-in; pass a 4-tuple (lo_d, hi_d, lo_r, hi_r) for any other wavelet.
• cwt: Gaussian 1st derivative continuous wavelet transform — the one used by ECG annotators. Other mother wavelets: pass pywt.

Scope

opendsp is intentionally narrow:

• ✓ Linear-time-invariant filter design + application
• ✓ Peak / extremum detection
• ✓ Wavelet decomposition (DWT + CWT)
• ✓ Resampling, smoothing, interpolation primitives

What it deliberately does not include:

• ✗ Spectral analysis higher than 1-D periodogram / FFT helpers (use scipy.signal.welch / numpy.fft directly)
• ✗ Adaptive filters, Kalman, ML-based denoising
• ✗ Modality-specific features (QRS, BSR, SEF95 etc. — those live in openecg, openeeg, openvital)

If you find yourself reaching for one of those, you probably want scipy.signal or one of the modality libs above.

License

Apache-2.0. See LICENSE.