Word-clip alignment: method and listening review
accent-atlas, https://mariuscomper.uk/accent-atlas/
Record dated 2026-09-22. Plain-text companion to the Sources list.

1. What the clips are
---------------------
Each speaker page plays six showcase words (Stella, store, snack,
brother, Wednesday, station) cut from that speaker's own paragraph
recording (Speech Accent Archive, CC BY-NC-SA 4.0). 126 clips in
all: assets/words/w-{speaker}-{word}.mp3.

2. How the boundaries were found
--------------------------------
whisper.cpp (small.en) token timestamps over each paragraph
recording, mapped to the paragraph words by difflib (alignment
ratios 0.80-0.96):
- 100 clips: direct or padded whisper token spans;
- 1 merged, 1 fuzzy-assigned, 3 interpolated (mid-word timestamps
  missing, neighbours interpolated);
- 21 "station" clips (the last word, where whisper end-timestamps
  overshoot or collapse): trailing-energy detection instead.
Every clip passed mechanical gates: duration 0.15-1.5 s,
mean volume above -50 dB. Boundary table: research/wordcuts.json.

3. Listening review (2026-09-22)
--------------------------------
Twelve clips whose alignment was uncertain were re-checked by
re-transcribing the cut region at several widths (whole clip,
1.0 s / 0.7 s slices at 0.4 s / 0.2 s steps across cut +/- 1.6 s)
with whisper.cpp small.en, plus 10 ms energy envelopes, and
confirming the target word sits inside the cut. Six passed as
cut; six were wrong and have been re-cut (new boundaries below,
files replaced in research/clips/ and shipped in assets/words/).

Passed as cut:
- 85/Stella [1.00,1.70] merged: clip transcribes "Steller".
- 141/store [3.08,3.58] fuzzy: clip transcribes "star" (York
  reading has no final r; consistent with the archive transcript).
- 11/station [19.52,20.47], 12/station [24.24,25.04],
  38/station [22.22,23.02], 110/station [25.55,26.20]: all
  transcribe "train station"/"station"; the word is fully
  present with a partial "train" lead-in.

Re-cut (old boundaries replaced):
- 24/brother -> [9.97,10.28] (old [10.46,10.74] held "Bob" and
  silence). New clip transcribes "Brother."
- 108/snack -> [9.39,9.64] (old [9.51,9.80] started mid-word).
- 108/brother -> [10.14,10.48] (old [10.35,10.62] held the word
  tail plus the "Bob" onset). Speaker reads "brother" with a
  GOAT vowel; both transcribers hear "roar/or".
- 57/station -> [24.37,24.88] (old [24.60,25.37] held the word
  tail plus trailing silence). New clip: "station."
- 113/station -> [22.37,22.86] (old [23.04,23.84] was silence
  after the speech ended). Context transcribes "station."
- 134/station -> [24.70,25.18] (old [25.12,25.92] was silence
  after the speech ended). New clip: "station."

All six re-cut clips re-pass the gates (0.25-0.51 s, mean
volume -16 to -29 dB) and re-verify by transcription. A random
sample of four direct-method clips outside the review list
(2/Stella, 40/Wednesday, 129/store, 145/snack) was context-
checked and confirms the direct alignment is sound.

4. Limits
---------
Cuts are word-centred, not word-exact: neighbours ("train",
"Bob") may bleed in at low energy on fast readings. Single
isolated words sometimes defeat the transcriber even when the
cut is correct (verified against wider context instead). The
IPA text on the page is the archive's own transcription and is
unaffected by clip boundaries.