Recreated acquaintance comparison
=================================

These commands apply early and later extraction rules to the SAME surviving
sentence. They do not regenerate speech and do not reproduce the missing
original take. The later cut matches the saved word audio sample for
sample. The early cut omits the opening burst of the final repetition.

Run from the listen-and-spell repository root. This uses a fresh output folder
and leaves the generated batch untouched.

AUDIO_SOURCE="experiment_01/generated/take3_en/audio/en/emily/195fc281cb.sentence.wav"
AUDIO_EXAMPLE_DIR="$(mktemp -d /tmp/acquaintance-example.XXXXXX)"

python3 experiment_01/extract_word.py \
  --min-silence 0.06 --tail-padding 1 \
  "$AUDIO_SOURCE" "$AUDIO_EXAMPLE_DIR/acquaintance-bad.wav"

python3 experiment_01/extract_word.py \
  --word acquaintance \
  --template "Please say the word '{WORD}.' {WORD}." \
  --estimate-rate --detect-sentence-start \
  --pause-selection nearest --pause-window 0.30 \
  --min-silence 0.06 --tail-padding 1 \
  --noise-db -40 --word-ms-per-letter 14 \
  "$AUDIO_SOURCE" "$AUDIO_EXAMPLE_DIR/acquaintance-good.wav"

Expected source SHA-256:
79c6354f84049bb0a81e8f097013c49cab84fa739ba3f94c76dab9bcb5af4ac3

Expected cuts:
  early rule: 2.755375 to 3.760000 seconds
  later rule: 2.489313 to 3.760000 seconds (rounded to source sample boundaries)

Source request:
  model: chatterbox
  voice: emily
  language: en
  input: Please say the word 'acquaintance.' Acquaintance.
  seed: 1234
  max_tokens: 1024
  options: exaggeration=0.5, guidance_scale=0.5, temperature=0.8,
           repetition_penalty=1.2

For the screenshots, open the downloadable acquaintance-sentence.wav in
audio-lab (npm --prefix experiment_01/audio-lab run dev). Load either
acquaintance-bad.selection.json or acquaintance-good.selection.json from the
blog. Use -35 dB and 0.06 seconds for the displayed silence bands in both
screenshots; these display settings are independent of the extraction rules.
The JSON fingerprints match the downloadable complete sentence.

The original screenshot UI rounds boundaries to milliseconds and may resample
for playback. The downloadable clips retain the original 24 kHz PCM samples.