Skip to content

0.1.0: first release

Latest

Choose a tag to compare

@kilickursat kilickursat released this 22 Oct 06:27

First release of WaveBench: signal processing in the browser for raw sensor records such as 3-axis accelerometer logs.

Open it: https://kilickursat.github.io/WaveBench/ — or download wavebench-offline.html below and open it from disk (one self-contained file, no installation, no network needed). Files are processed on your device and never uploaded.

What it does

  • Import CSV, TSV and TXT: delimiter, decimal comma, header and metadata lines detected and adjustable; missing or unreadable cells kept as missing, with line numbers.
  • Timebase: enter the sampling rate, or check a time column (s, ms, µs, ns or ISO 8601) for order, gaps and jitter, with the rate fitted to the timestamps. Irregular records are refused for FFT work until you choose a gap-free range or an explicit rate.
  • Units and calibration: offset and gain per channel; exact conversions between g, mg, m/s², mm/s² and Gal; optional derived |a|.
  • Filters: Butterworth low-pass, high-pass, band-pass, band-stop and notch, as second-order sections, zero-phase or causal; the response of what is applied, and start-up regions marked on the time plot.
  • Analyses: FFT (peak, RMS, power, PSD, ASD; any length, five windows), Welch PSD with band RMS, STFT spectrograms with linked cursors, statistics, the data table.
  • Exports: CSV of every result, filter coefficients, PNG figures, and a recipe (JSON) with every setting and the input's SHA-256.
  • Compute: Float64 on the CPU; WebGPU (Float32) for spectrograms by default and for FFT and Welch on request, with a CPU fallback and an in-app check of your GPU against the CPU.

Validation

  • CPU engine against NumPy 2.3.5 / SciPy 1.17.0 fixtures (FFT, windows, detrending, periodogram, Welch, STFT, Butterworth and notch) and 50-digit mpmath references for filters: 250 unit tests.
  • WebGPU against the Float64 engine and the SciPy fixtures: 31/31 comparisons within 1e-4, measured errors about 1e-7 (on the SwiftShader software adapter in CI).
  • Details: docs/VALIDATION.md.

Known limits

  • WebGPU has been validated on a software adapter only; use Compute → Check this GPU against the CPU on your hardware.
  • GPU transforms use power-of-two FFT lengths; other lengths run on the CPU (stated with each result).
  • No resampling, cross-spectra or integration yet — see the roadmap.