Help
Toolbar
WAV Files: Select WAV files. All decodable WAV files are accepted; filenames are used only for metadata. Categories are inferred from filenames of the form <categoryToken>_<anything>.wav or <categoryToken>-<anything>.wav and displayed as compact 1- or 2-character codes.
◀ PREV: Return to the previous file.
NEXT ▶: Process/display the next file.
File chip: Shows the current file number, e.g. 3/10.
EXP: Export ZIP.
📊 SCATTER: Scatter plots.
◀ BAK / FWD ▶ / ALL: Zoom nav.
↺ RST: Factory reset.
F0 display
F0: Switches the time canvas from the standard waveform + envelope display to a waveform hint + F0 display. In F0 mode the button is labelled ENV and switches back to the standard display.
The F0 procedure uses a DC-centred and normalised waveform, 1 kHz downsampling, a user-facing 2nd-order Butterworth F0 band-pass prefilter, Hirst-DeLooze F0 limit estimation, YIN F0 estimation, linear bridging of internal unvoiced sections, and a moving median F0 window. Originally voiced F0 is drawn solid blue; originally unvoiced linearly bridged intervals are lighter blue and dotted.
Frequency domain
Envelope-spectrum peaks are rhythm-formant candidates. Fuzzy shaded bands indicate approximate RFA reference zones around syllable-pair, syllable-rate, and fast-syllable edge regions.
The default RFA range is 1–6 Hz with conservative DC removal at 0.1 Hz. Shaded bands are RFA reference zones, not measured evidence. The peak-zone curve is kernel-smoothed evidence from spectral peaks. Candidate rhythm-formant zones combine peak clustering, prior overlap, and temporal plausibility; they are candidates only, not final rhythm identifications.
The LF spectrum shows the raw low-frequency envelope-spectrum magnitudes. The RFA-smoothed LF spectrum suppresses narrow bin-to-bin fluctuations and highlights broader rhythm-formant regions. Candidate RFA peaks and zones should be interpreted with this smoothing relationship in mind: the smoothed curve clarifies the LF spectrum but does not constitute a second independent measurement.
MOVWIN: MA uses an equal-weight moving average, GAU uses Gaussian weights, HAN uses a Hann taper, and TUK uses a Tukey taper with selectable α. These methods smooth neighbouring frequency bins only; the envelope low-pass cutoff is a separate time-domain smoothing control.
MOVWIN bins: selects the moving-window length in spectrum bins: 3, 5, 7, 9, 11, 13, or 15. If the spectrum of the selected envelope has fewer than twice the selected MOVWIN bin length, MOVWIN is not applied and a warning is shown.
nPVI measures
IPUnPVI: interpausal-unit nPVI; excludes intervals that cross silences/pauses and is the default statistic for scatter and spreadsheet calculations.
GLOBnPVI: global nPVI; includes all consecutive peak intervals and is retained as a diagnostic comparison.
Scatter machine learning
Plain: category-coloured scatter with file-number point labels.
k-means: unsupervised clustering in Duration, Rate, and selected nPVI. Colours show K-clusters; labels show category plus file number.
Linear SVM: binary supervised boundary on the displayed two-feature scatter axes. The solid line is the decision boundary, dotted lines are margins, and marked points are support vectors.
Kernel SVM: binary RBF-kernel SVM on the displayed two-feature scatter axes. The solid contour is the non-linear decision boundary; dotted contours are approximate margin contours. C and gamma are compact controls. Fit agreement is descriptive and is computed on displayed training data.
DENDRO: three one-variable horizontal dendrograms for Duration, Rate, and the selected nPVI measure. DENDRO mode is exploratory teaching/study output and does not validate category labels or k-means clusters independently. For one-variable dendrograms, Manhattan and Euclidean distances both reduce to absolute difference; linkage choice usually has the larger visible effect. Ward linkage is provided only with Euclidean distance.
For SVM target K-means: The SVM - k-means comparison is exploratory only and does not imply validation.