Skip to contents

Changelog

Source: NEWS.md

shinytone (development version)

  • Model: GAMM tab keeps a <tier>_tseq axis on its own scale. For multisyllabic words the guide routes you to fit on syllable_tseq (syllable 1 spans 0–1, syllable 2 spans 1–2), but fit_gamm() rescaled each token’s time to 0–1 like any other column. That squeezed every token’s first-to-last frame onto [0, 1], so the syllable boundary no longer sat at the same point across tokens, and the plot showed a 0–1 axis. fit_gamm() now recognises a _tseq column and uses it as-is (time_normalised = "sequential" declares any such axis). predict_gamm() predicts across its observed range, and the fit records time_axis, time_range and n_segments. The GAMM plot keeps the 0-to-n axis, marks each segment boundary with a dashed line, and labels the segments. The tab explains the axis when _tseq is chosen, and Show R code reproduces all of it. Summarise now declines Chao numerals for such a fit, since they describe one syllable on a 0–1 axis.
  • Axis titles: f₀ (st) for semitones. The GAMM plot labelled its y axis with the raw column name (Predicted f0_st); it now uses the same title as the other tabs (Predicted f₀ (st)). Semitone axes read f₀ (st) rather than f₀ (semitone) throughout: Visualise, Curate, Cluster, GCA and GAMM.
  • F0 Analysis Start tab: add a contour duration column. A new Add contour duration block under Attach metadata adds how long each f0 contour lasts: the time from the first to the last frame with f0, per token (token_f0_dur) or per landmark segment such as each syllable (syllable_f0_dur), repeated on every row of the unit. Unvoiced frames at the edges do not count, and unvoiced frames inside the contour do not shorten it. The ends are anchored on two voiced frames in a row, the rule the Whole token export region uses, so a lone stray frame in the silence cannot stretch the duration; a whole token’s value then equals the voiced_s that F0 Extraction reports. The column reaches every F0 Analysis tab. Also scriptable as contour_duration() (min_run = 1 for the literal first-to-last frame), which counts negative values as measured, so it works on semitone or z-score columns too.
  • F0 Extraction tab: Drop rows without f0 export option. The Whole token region already leaves out leading and trailing silence but keeps unvoiced gaps inside the contour as f0 = NA rows. A new checkbox (off by default) drops those rows too, for tools that expect measured values only. It runs last, so resampling still sees the gaps and has_gap / n_missing still flag the tokens. With equidistant points, a token with a gap then keeps fewer than N points, each still labelled by point and time_prop.
  • F0 Extraction tab: 0 Hz in an uploaded f0 CSV reads as unvoiced. The wrassp and .Pitch paths already turned 0 Hz into NA, but a CSV’s zeros passed through as values and could reach the export inside the region.
  • F0 Processing: edit history survives the F0 Data Export round trip. The export writes corrected values into f0 but recorded nothing about which frames were corrected. Re-uploading it therefore loaded the corrected values as the tracker’s originals, and the next all_correctedf0.csv read edited = FALSE everywhere. When corrections exist, the export now also writes f0_original (the tracker’s value) and edited, and re-uploading either file restores the edits. This works whichever f0 column the picker is set to: choosing f0_corrected from all_correctedf0.csv used to lose the history the same way. Re-uploading an older export without these columns now shows a note explaining that its corrections cannot be told apart. The export also stops copying correction columns from a re-uploaded file: a token restored in a later session kept a stale token_dropped = TRUE, so F0 Analysis silently excluded it.
  • F0 Correction tab: downloads carry landmark and metadata columns. all_correctedf0.csv and <token>_f0.csv held only the f0 frames plus f0_corrected / edited / token_dropped, so the metadata (uploaded, or derived from filenames) and TextGrid landmark columns set up in F0 Extraction were missing after correction. Both downloads now add them the same way the F0 Data Export does. Re-uploading all_correctedf0.csv to resume keeps these columns, and attaching the same metadata again does not duplicate them: a metadata column the data already holds with the same values is skipped. This also drops the redundant token.meta column that filename-derived metadata used to add.
  • F0 Extraction tab: landmark tiers ticked after extraction reach the export. Landmark columns were attached only when Run extraction was clicked, and the Landmarks from TextGrid picker sits below that button, so ticking syllable afterwards (or in Upload existing f0 CSV mode) left the download without syllable_start / syllable_end / syllable_i. The Normalise tab lists a tier only when those columns exist, so multisyllabic data had no tier to build <tier>_tseq from. The export now attaches any ticked tier the data lacks. An amber note also warns when Equidistant points drops the landmark columns (their boundaries change within a multisyllabic token), and the Normalise tab explains why no tier is listed.
  • F0 Correction tab: two fixes from classroom use. Deleting an outlying frame could leave its “ghost” marker (the original value) drawn off-screen: the f0 panel’s y-range was framed on the corrected values only, and the y-axis is fixed, so there was no way to scroll the original back into view. The range now also anchors on the original values of edited frames. And the Delete button sometimes appeared to do nothing: a click after a box/lasso selection silently kept the stale box (plotly never clears that input on a plain click), so the edit hit the old — often already deleted — frames. The selection now follows whichever of click or box-select happened last, and switching tokens clears it.
  • F0 Correction tab: a note column in the edit log. A Note (optional) box in the sidebar. Whatever it holds is saved with the next edit or discard as the note column of the edit log — table, CSV download, and log re-upload alike — then cleared so it cannot leak onto later, unrelated actions. For recording why: a reason to discard a token, or why frames were removed.
  • F0 Extraction tab: export notes restyled as a banner. The data-quality notes under the export summary (short voiced span, unvoiced gaps, …) were easy to miss as small grey text at the bottom of the sidebar; they are now an amber message box in the same idiom as the Correction tab’s banners.
  • Praat script: version guard for the newer pitch methods. The filtered ac / filtered cc / raw ac / raw cc methods exist only since Praat 6.4 (November 2023); on an older Praat the batch died mid-run with a cryptic “Command not available for current selection” error. The script now checks praatVersion up front and exits with the running version and the fix (update Praat, or pick ac / cc / shs).
  • F0 Extraction tab: F0 Data Export (region + sampling). A new block decides what the downloaded dataset contains, without touching the data in the app: F0 Correction keeps working on the extraction exactly as it arrived, so changing these settings can never disturb frame edits. Region is what the measurement covers — Whole token runs from a token’s first to its last voiced frame, so leading and trailing silence (which carries no f0) is excluded, and an edge needs two consecutive voiced frames so a lone voiced frame stranded in silence cannot stretch it; or a TextGrid interval. Sampling is how densely — every native frame time, or N equidistant points across the region, which is time normalisation applied at the sampling stage and adds point (1…N) and time_prop (0–1). An export preview shows the first rows of the file the Download button will write, metadata join included. New exported functions: resample_f0_equal(), trim_to_voiced(), flag_f0_gaps().
  • Resampling follows Praat’s own rule. resample_f0_equal() ports Sampled_getValueAtX(), which is what Pitch: Get value at time... runs, so a contour resampled here matches one a Praat script gets querying the same times. Of the two frames bracketing a point, the nearer is near and the other far: both voiced blends them, far unvoiced keeps near’s measured value, near unvoiced gives NA. The method argument offers Praat’s two choices, "linear" (default) and "nearest". Unvoiced stretches are respected only when the input marks them as NA (or 0 Hz) rows; sparse input such as a .PitchTier has no such rows, so a silent stretch there is a plain gap between anchors and is interpolated across, exactly as Praat does.
  • F0 Extraction tab: subset f0 by TextGrid interval. Export only the f0 falling inside chosen intervals: vowels found automatically, the rhyme (first vowel to the end of the token; monosyllables only), or labels you type. New exported ipa_vowel_label() and filter_interval_rows(). ipa_vowel_label() handles length marks, stress marks, tone digits and Chao letters, combining diacritics, precomposed pinyin tone marks (ā ǎ ū ǔ …) in either Unicode normalisation, di- and triphthongs including j/w/ɥ offglide spellings (aj, ɔw) as well as ai and au, and syllabic nasals (m̩, n̩, ŋ̩) as vowel-equivalent nuclei.
  • Two checks that make silent problems visible. flag_f0_gaps() adds n_missing and has_gap, marking tokens whose voicing was interrupted mid-region — measured on the native frames before resampling, since resampling can fill a short dropout from the nearer frame. And the per-token summary reports voiced_s with voiced_prop (voiced_s / duration_s), shading and naming tokens voiced across a far shorter span than the rest of the corpus: with equidistant sampling their percentage positions are squeezed into whatever was tracked, so they are not comparable with the other tokens.
  • Modelling: an already-normalised time column is used as-is. fit_gca(), fit_gamm(), fit_polynomial() and compute_mean_contour() gained time_normalised ("auto", "no", "yes"). Under the default "auto", a column that is already proportional — token_t01 from normalise_time_token(), or time_prop from equal-point extraction — is detected by the new time_already_normalised() and used as-is, rather than min-max rescaled a second time, which stretched any token not spanning the full unit interval. The Model, GCA and GAMM tabs say when this applies and their Show R code output reflects whichever branch ran; fit_gca() and fit_gamm() record time_prenormalised in the returned object, fit_polynomial() as an attribute. Detection is deliberately conservative and has two documented limits: a set in which every token is a partial span cannot be distinguished from ordinary variable-duration time and is rescaled as before (pass "yes"), and raw seconds in which every token lasts just under one second is detected as proportional (pass "no").
  • 0 Hz counts as unvoiced in the new f0 helpers, matching inspect_f0() and the extraction paths, so trackers that code unvoiced frames as 0 Hz are handled rather than having silence blended into speech.
  • The multisyllabic warning no longer fires on token_t01. On the Model, GCA and GAMM tabs it is now raised only for a landmark tier’s <tier>_t01, which resets at every segment boundary, and not for the whole-token proportional axis, which is a perfectly good time variable.
  • F0 Correction tab: whole-token discard. A new “Whole token” edit group adds Discard token / Restore token for tokens that are beyond repair: instead of fixing frames, the whole token is marked as dropped. Non-destructive, since the f0 values are kept and both downloads gain a token_dropped column (TRUE for discarded tokens) to filter on downstream. Discarded tokens show a ✗ in the token picker and a banner above the plot, appear in the edit log (Undo restores, as does the Restore button), survive the save/re-upload resume cycle via the new column, and a “Kept + discarded / Only kept / Only discarded” filter joins the edit-status drawer. Keyboard: X discards the current token, or restores it if already discarded, so a review pass can run entirely on , . and X. The sidebar progress line and every discard notification report the running share of the corpus discarded, e.g. “Discarded: 812 of 8000 (10.2%)”.
  • F0 Correction tab: bulk discard of flagged tokens. Once an Inspect-tab CSV is loaded in the filter drawer, a Discard all flagged tokens button marks the entire flagged set as discarded in one click. The confirmation dialog reports how many tokens that is and what share of the corpus they represent. This is the fast route for large corpora: discard the flagged set, then optionally review it (“Only discarded”) and Restore any worth repairing. A Restore all button next to the discard toggle un-discards every discarded token, bulk and manual alike, and removes their edit-log rows.
  • F0 Correction tab: discard-share breakdown by speaker and tone. The corpus-wide discarded share can look harmless while the discards pile up in one speaker or tone — worst case, a whole speaker × tone cell is emptied and vanishes from the retained data. When the filter drawer’s speaker / tone columns are picked (auto-guessed from the uploaded Inspect CSV), the bulk-discard confirmation dialog shows what the discard set would look like per speaker × tone cell with marginals (“After this discard”), and a Breakdown button next to the discard toggle reopens the same table for the current state. Cells are discarded/total (%), tinted red when a group would be fully discarded and amber at half or more, with tokens absent from the CSV grouped as “(no metadata)”. With only one of the two columns picked the table reduces to that margin; with neither, the dialog shows the total plus a hint to pick them.
  • GCA and GAMM guides: pointer to model structures the UI does not cover. Both guide boxes now close with a Beyond the built-in options note. The checkboxes and dropdowns cover the structures most often used for tone contours, but lme4 and mgcv support many more — by-speaker tone slopes (1 + (ot1 + ot2) * tone | speaker) or uncorrelated terms via || for GCA; tensor-product interactions such as te(time_norm, duration), or smooths varying by a further factor, for GAMMs. The note directs users to Show R code, which already emits a runnable script reproducing the current fit, as the starting point for editing the model formula.
  • Model: Polynomials — Show R code now reproduces the coefficient scatter. The tab’s generated script used to stop at the coefficient table, leaving the coefficient-space scatter as the one part of the tab with no reproducible counterpart outside the app. The snippet now ends with a plotting section that rebuilds the scatter — one point per token, coloured by tone — using the axes currently selected in the tab (falling back to the plot’s own defaults) and axis labels carrying the phonetic gloss (c1 slope, c2 curvature, …). With no Z axis it emits a static ggplot2 scatter; with a Z axis selected it emits the matching 3-D plotly::plot_ly(type = "scatter3d") call instead, mirroring what the tab shows.
  • F0 Correction tab: filter by flag type. When the uploaded Inspect CSV carries flag_notes, a Keep flag types checkbox group lists the artefact classes present (extreme value, level, octave jump, jump by rate of change, carryover, low intensity). Unticking a type hides tokens that carry none of the ticked ones and narrows Discard all flagged tokens to the same subset, so a corpus can be worked one artefact class at a time (e.g. discard the octave jumps, review the level outliers by hand). It combines with the discard-status filter, so the discarded set can be reviewed one flag type at a time.
  • F0 Extraction: Praat is the default f0 source when pitch files are uploaded. Uploading .Pitch / .PitchTier files alongside the audio now selects “Use uploaded .Pitch / .PitchTier (Praat)” automatically and says so, instead of leaving the radio on wrassp and silently extracting without Praat’s per-frame candidate lists. It fires once per session, so a later manual choice is never overridden.
  • F0 Correction: Praat candidates promoted to an edit group. Picking a candidate writes to the contour, pushes undo history and logs an edit row, so the block now sits with the other edit groups (after Manual entry) rather than below the Display checkboxes. When no .Pitch data is loaded it shows a short hint explaining how to enable the option, so the feature is discoverable from a .wav-only session. The candidate list now spells out that s is Praat’s strength and that the tick marks the frame’s current value, and the empty state mentions that the grey numbered dots on the plot can be clicked directly. The Display checkbox “Top-3 Praat candidates on f0 plot” is greyed out and unticked when the data carries no candidates, rather than sitting ticked over an overlay that cannot appear, and its caption explains that dot 1 is the value Praat chose (2 and 3 are its next alternatives), that not every frame has three candidates, and that the unvoiced candidate is not numbered.
  • F0 Correction: plot mark legend. A “Marks:” strip above the plot names every non-obvious mark: f0 value, selected frame, Praat candidates (dot 1 being Praat’s own pick, click to apply), the two sample-level flags in parallel wording (“jump or carryover”, red fill; “low intensity”, a bare amber ring that can sit on any fill colour), edited frames, and the outlined circle showing a frame’s value before the edit. Entries appear only when the corresponding data exists. When an Inspect CSV is loaded, a tinted status box under the key summarises the current token: light coral (dot-red border) whenever the token is flagged, with the frame counts and flag classes, or, for a token-level flag with no flagged frames (extreme value / level), a pointer to look at the whole contour; amber when the only signal is the advisory low-intensity ring; green with an explicit “nothing flagged by Inspect”. Flagged states add the reminder “Flags are leads, not errors: verify by eye and ear before editing” on its own line.
  • F0 Correction: low-intensity frames marked on the plot. Frames the Inspect tab flagged as low intensity now carry an amber marker ring (reading flag_low_intensity when present, else the flag_notes text). Kept deliberately distinct from the red fill: red means a probable tracking error, the amber ring only means the f0 estimate there is less reliable. The ring co-exists with the red/blue fills, hover text says “low intensity (f0 here is less reliable)”, the legend names it “flagged by low intensity”, and the “Keep flag types” filter’s Low intensity class now has a visible counterpart on the plot.
  • F0 Correction: re-extraction no longer strands earlier edits. After re-extracting on a different engine or frame grid (e.g. Praat first, then wrassp), edits made on the old grid were kept keyed by token name: the tab silently served the stale contour, and because the row counts no longer matched, ghost markers, the amber halo, and the “N frames edited” banner all vanished for that token while edits still appeared to apply. Frame edits that no longer align with the data are now cleared on re-extraction (with a notification); whole-token discards, which are name-keyed and still meaningful, survive.
  • F0 Correction: the idle reminder dismisses itself. The “Still working?” note that appears after ten idle minutes used to stay until closed by hand; the first interaction after it fires now takes it down (it returns after the next ten idle minutes).
  • Curate tab: “Flagged” now covers every Inspect check. The amber highlight, the “Flagged” quick-select, and the flagged-count chip now use flagged_token (any check: extreme max/min, unusual level, frame-level jumps) instead of only the “level too high / low” notes, so the exclude machinery can also serve as a whole-token disposal path for artefact-flagged tokens.
  • Start tab: discarded tokens honoured. Uploading a CSV that carries the F0 Correction tab’s token_dropped column (e.g. all_correctedf0.csv) now pops a notification reporting how many tokens are marked discarded and excludes those rows (and the flag column) from the working dataset by default; a sidebar checkbox (“Exclude discarded tokens”) restores them. The uploaded file itself is never modified. Previously the discarded tokens flowed silently into every downstream tab (Normalise, Inspect, Visualise, the models, Summarise).
  • Bulk-discard guidance. An illustrated “Big corpus, small flagged set? Bulk discard and review” guide joins the F0 Correction tab, walking through flagging in Inspect, discarding the set, and the optional review pass. When a corpus is large but lightly flagged (at least 1000 tokens with at most 15% flagged; thresholds in offer_bulk_triage()), the Inspect summary and the F0 Correction flagged-CSV loader suggest that route proactively.
  • flag_outliers() (the speaker-level extreme-value screen) is now one-sided: a token is flagged too_high only when its per-token maximum is unusually high (z_max > z_threshold), and too_low only when its minimum is unusually low (z_min < -z_threshold), replacing the previous two-sided abs(z) > z_threshold. Gross tracking errors are directional (octave-doubling or a spurious spike inflates the maximum; octave-halving, a subharmonic, or creak deflates the minimum), and because the screen pools all of a speaker’s tones the opposite tails hold legitimate low/high tones rather than errors. This makes the flag_too_high / flag_too_low columns — and the “max too high” / “min too low” notes from inspect_f0() — directionally correct; before, a token with an unusually high floor could be mislabelled “too low”. The set of flagged tokens is essentially unchanged on clean data (the rarer truncated-max / floored-min cases are left to flag_level_outliers() and flag_pitch_jumps()).
  • inspect_f0() now accepts tone = NULL, which skips the tone-relative token-level check (flag_level_outliers()) and omits the tone column from the output, so it can run before tone categories are known (e.g. the clustering / tone-discovery workflow). The speaker-level extreme-value and sample-level jump checks still run, and the default remains tone = "tone", so existing calls are unchanged.
  • Inspect tab: optional tone. The tone selector now offers — none —, which runs the two tone-free screens (speaker-level extreme-value and sample-level jumps) without a tone column, for the pre-tone-discovery workflow. Backed by inspect_f0(tone = NULL).
  • GAMM tab: on-demand diagnostics. The “Run model diagnostics” button is shown in the sidebar from the start (disabled until a model is fitted), and diagnostics run when it is clicked rather than automatically after every fit, so fitting stays fast.
  • GAMM diagnostics (new). diagnose_gamm() and a “Model diagnostics” section on the GAMM tab: after fitting, one click reports the basis-dimension check (k’, edf, k-index, p-value, flagging under-resourced time smooths), the four gam.check() residual panels (Q–Q, residuals vs fitted, histogram, observed vs fitted), the residual ACF, and the concurvity table, with a text download of all diagnostics. The ACF is computed per token (in the spirit of itsadug::acf_resid()) and AR1-whitened when the fit used an AR1 correction, so it shows whether the correction actually worked.
  • GCA fit-over-data overlay (new). The GCA tab now optionally overlays the observed per-tone mean contour (semi-transparent points, computed with compute_mean_contour()) on the fitted polynomial curves, so you can judge how well the chosen degree tracks the data — the standard GCA model-adequacy check (Mirman 2014). Toggle it with the “Overlay observed per-tone means” checkbox; turn it off for a cleaner plot on busy data. The overlay is included in the downloaded plot.
  • The GAMM tab’s AR1 correction is now on by default: densely-sampled f0 frames are strongly autocorrelated, so smooth p-values are anticonservative without it. fit_gamm() (whose use_ar1 argument still defaults to FALSE) now estimates rho from the lag-1 autocorrelation within tokens rather than across the flat concatenation, so token boundaries no longer bias the estimate.
  • run_app() now checks GitHub once per launch for a newer shinytone release (2-second timeout, silent when offline) and, when one exists, prints the update command in the console and shows a one-time notification in the app. Disable with options(shinytone.check_updates = FALSE).

shinytone 1.0.0

First stable release. The Shiny app covers the citation-tone workflow end to end (collect, extract, inspect, correct, curate, normalise, cluster, visualise, model, Chao numerals), and the analytical functions behind each step are exported and documented. Highlights since the 0.1.x line:

  • Contour clustering (new). An unsupervised “tone discovery” workflow that groups tokens by f0-contour shape when the number of tone categories is unknown (Kaland 2023). New functions: cluster_f0() (k-means, hierarchical Ward, or Gaussian-mixture clustering), cluster_features() (represent each contour as resampled points, Legendre / DCT coefficients, or its derivative), choose_k_f0() (suggest the number of groups via silhouette, gap statistic, and a minimum-description-length cost), cluster_mdl(), and cluster_agreement() (adjusted Rand index against provisional labels). Surfaced through the new Cluster tab.
  • Contour sonification (new). sonify_f0() renders an f0 contour as an audible waveform: a pure tone, a complex tone (12 harmonics, band-limited below Nyquist), or a source-filter synthesised vowel (a / i / u), with the pitch gliding along the contour. An optional intensity argument shapes the loudness envelope from per-frame dB. The Cluster tab’s “Listen to the contours” panel plays each candidate cluster’s mean contour back, so prototypical tones can be heard, not only seen (faithful Hz when an Hz column is present, or shape-only on a chosen base pitch).
  • Hear your corrections (new). The F0 Correction tab sonifies the contour you are editing as an Extracted vs Corrected A/B pair, played at the token’s own duration with loudness following the measured intensity, so you can hear whether an edit fixed the pitch track.
  • Intensity made visible (new). In F0 Correction, the f0 dots are sized by per-frame intensity (louder = bigger), with a size legend and per-frame dB on hover.
  • Curate (new). apply_relabels() re-labels tone-category variants (splits or mergers, colloquial vs. literary readings, sandhi) or excludes mis-elicited tokens without overwriting the original labels, surfaced through the new Curate tab.
  • TextGrid landmarks (new). When Praat .TextGrid files are supplied, the F0 Extraction step can attach per-frame landmark columns from a chosen interval tier: <tier>, <tier>_start, <tier>_end, and <tier>_i (segment index). New functions: tg_interval_tiers(), assign_tier_landmarks(), and attach_landmarks(). The Visualise tab can then align contours by these landmarks, including syllable by syllable for multisyllabic words.
  • Landmark time normalisation (new). normalise_time_landmarks() rescales time within each segment, adding a within-segment 0–1 axis (<tier>_t01) and a sequential, word-level axis (<tier>_tseq). A normalise_time_token() companion rescales the whole token to 0–1 (no landmarks needed, for monosyllabic data). Surfaced through a new “Time Normalisation” section on the Normalise tab. The model tabs default the Time variable to <tier>_tseq when present and steer multisyllabic analyses toward GAMM.
  • F0 Processing. The Start preview now flags audio files too short to yield an f0 frame and skips them during extraction. flag_low_intensity() (the intensity-based inspection check) is now exported and documented.
  • Visualise palette (new). The tone colour scale is a 12-colour set matched to the app theme, replacing the pale default so every tone reads clearly on a white background.
  • Faster start-up (new). A one-time “loading analysis tools” toast appears the instant the page connects while the heavy analysis packages load in the background, so the landing page paints immediately.
  • Praat extraction script (new). The bundled script writes readable text .Pitch files, samples per-frame intensity into the CSV, and resolves the chosen CSV output path correctly on Windows.

shinytone 0.1.2

  • New flag_level_outliers(): a third inspection layer that compares each token’s overall level (its median f0, in semitones) against other tokens of the same speaker and same tone using a robust modified z-score (median/MAD; Iglewicz & Hoaglin 1993, cutoff 3.5). It flags smoothly shifted contours — e.g. a low-tone token mis-tracked up into the mid-tone band — that the pooled max/min check and the sample-level jump check both miss. Surfaced through inspect_f0() (new level_threshold and min_tokens arguments, level too high / level too low notes) and the Inspect tab.
  • The "norm" (normalised-time) option was removed from time_unit in flag_pitch_jumps(), inspect_f0(), and the Inspect tab: the rate-of- change thresholds are physiological (ST per 10 ms) and have no meaning once real time is discarded. Inspection runs on real-time data (s / ms).
  • Inspect-tab guide rewritten around three complementary layers, and the “Pitch-tracking quality check” workflow now marks Normalise as optional (inspection runs on raw f0).

shinytone 0.1.1

  • flag_pitch_jumps() (and therefore inspect_f0() and the Inspect tab): when a sample-to-sample jump is detected, the flag is now placed on whichever side of the jump is farther from the token’s median f0, rather than always on the landing sample. This correctly identifies the artefact whether it sits at the start or the end of a sequence (e.g. an octave doubling on the first frame of a token, which the previous landing-only logic mis-flagged).
  • Carryover now walks both forward AND backward from each flagged sample, so artefact runs that begin or end the token are extended in both directions.
  • Inspect-tab guide text and flag_pitch_jumps()’s function docs describe the new median-aware logic as an adaptation of the rate-of-change + carryover approach in Steffman & Cole (2022).

shinytone 0.1.0