Testing

Running the test suite

# Run all 36 tests (no ONNX model required)
cd precise-onnx-js
node --test test/*.test.js
# or: npm test

Requires Node.js 18+. No npm install is needed. The library has no runtime dependencies beyond an ONNX runtime, which is only needed for load(). Tests mock the session.

Regenerating Python reference vectors

# Requires sonopy and ovos-ww-plugin-precise-onnx in the Python venv
"/home/miro/PycharmProjects/HiveMind Workspace/.venv/bin/python" \
    test/generate_ww_vectors.py

Writes test/ww_vectors.json. Commit the updated file so JS test runs are reproducible without a Python environment.

---

Test coverage (36 tests)

ThresholdDecoder (6 tests)

Test	What it checks
Constructor properties	minOut, maxOut, outRange, center, CDF length all match Python
CDF first 5 values	Floating-point match to 1e-12 tolerance
CDF last 5 values	Floating-point match to 1e-12 tolerance
decode on 13 values	Each decoded value matches Python to 1e-6
Boundary: decode(0) → 0	Exact
Boundary: decode(1) → 1	Exact

TriggerDetector (7 tests)

Test	What it checks
Seq 1 results match Python	Fire pattern: [F,F,F,T,F,F,F,F,F,F,T], firing on the 4th consecutive
Seq 1 activations match Python	Internal counter after each step
Seq 2 (all zeros) never fires	10 steps of 0.0 → all false
Seq 3 (boundary 0.5) never fires	prob > 1-sensitivity is strict, and 0.5 is not > 0.5
Cooldown value matches Python	-floor(8 * 2048 / chunkSize) = -8
Re-fires after cooldown	8+ low-prob steps + 4 high → fires again

PreciseOnnxWakeWord properties (5 tests)

Test	What it checks
windowSamples	1600
hopSamples	800
bufferSamples	24000
nFeatures	29
nMfcc	13

All values compared against precise_props from ww_vectors.json.

_updateVectors: rolling buffer (5 tests)

Test	What it checks
Zero-length chunk returns zero matrix	No frames computed
Short chunk (< windowSamples) accumulates	No frames yet, audio stored
windowSamples chunk → 1 frame	Exactly one MFCC row produced
After consumption, buffer drains correctly	windowAudio.length === hopSamples after 1 frame
MFCC matrix rolls on second chunk	Oldest rows drop, new rows appended

mfccSpec equivalence: zeros (4 tests)

Test	What it checks
Frame count matches Python	n_frames
frame[0][0] matches Python	Log-energy value (float32 precision)
frame[0][1..12] all zero	DCT of zero log-mels is zero
ATOL for all coefficients ≤ 1e-3	All 13 coefficients, all frames

mfccSpec equivalence: 440 Hz sine (4 tests)

Test	What it checks
Frame count matches Python
frame[0] all coefficients within ATOL	1e-3
frame[-1] all coefficients within ATOL	Last frame
No NaN/Inf in any frame	Numerical stability check

mfccSpec equivalence: white noise seed 42 (4 tests)

Test	What it checks
Frame count matches Python
frame[0] all coefficients within ATOL	1e-3
frame[-1] all coefficients within ATOL
No NaN/Inf

Integration: predict pipeline (1 test)

Test	What it checks
predict with mock session returning 0.99	Fires after 4 consecutive calls and returns boolean

---

Cross-language validation methodology

1. test/generate_ww_vectors.py runs with the Python reference (sonopy, ovos-ww-plugin-precise-onnx) and writes exact float32 values to test/ww_vectors.json.
2. test/wakeword.test.js loads the JSON and compares JS output against those values.
3. MFCC coefficients use ATOL = 1e-3 to account for float32 truncation. Python stores np.float32, and intermediate calculations are float64 in both Python and JS.
4. Probability/CDF comparisons use ATOL = 1e-6 or 1e-12 since those remain float64 throughout.

---

Test vectors (test/ww_vectors.json)

Key	Contents
threshold_decoder	min_out, max_out, out_range, center, cd_len, cd_first5, cd_last5, decode map
trigger_detector	cooldown, seq1 (fire pattern + activations), seq2, seq3
mfcc.zeros	Frames for 4800-sample zero audio
mfcc.sine_440	Frames for 6400-sample 440 Hz sine
mfcc.noise_seed42	Frames for 6400-sample uniform noise (seed 42)
precise_props	window_samples, hop_samples, buffer_samples, n_features, n_mfcc

---

← Wake word API · Home · AI usage →