Files
hpr-intro-stripper/tests/test_audio_correlation.py
Lee Hanken fcfa1f1c6c Fix gap-bridging correlation, a marker-normalization bug, and add sponsor ad-read markers
- audio.py: last_high_correlation_run_seconds() replaces a simple
  leading-run count for the theme.flac correlation specifically. The TTS
  speech is mixed additively onto the theme bed (sox -m), so correlation
  predictably drops once speech masks it even though the theme is still
  playing underneath; the old algorithm couldn't recover once correlation
  legitimately resumed afterwards (a genuine silence/instrumental gap
  before the host's own content). Requires the first qualifying run to
  start near the beginning, so a later island of coincidental correlation
  is never trusted without a confirmed anchor. Verified live: hpr4775 now
  resolves to 40.5s, matching the real boundary (confirmed by waveform
  inspection) almost exactly, instead of 27.3s.
- markers.py: fixed fuzzy_contains() comparing a raw (un-normalized)
  marker string against normalized text -- a marker containing
  punctuation, like 'archive.org', could never match. Found by a failing
  test, not by inspection.
- markers.json (version 4): enumerated the two sponsor ad-read scripts
  observed so far (anhonesthost.com, archive.org/Internet Archive) as
  markers, the same way outro_markers are -- not a generic 'small gap
  after a match still counts' heuristic, which was considered and
  rejected since it would swallow a host's own content whenever they talk
  immediately after the TTS block with no gap, which is the common case.
- docs/DESIGN.md: marked the silence-gap limitation resolved, documented
  the normalize bug, and added a short discussion of why general
  music/audio detection wouldn't have solved the sponsor-ad-read problem
  (it's audio-feature-identical to host speech) even though it was and
  remains the right tool for the music-bed boundary specifically.
- New regression tests for the correlation fix and both marker additions.
2026-11-22 11:51:46 +00:00

44 lines
2.0 KiB
Python

import unittest
from hpr_stripper.audio import last_high_correlation_run_seconds
class TestLastHighCorrelationRun(unittest.TestCase):
def test_bridges_a_masked_gap_like_hpr4775(self):
# Modelled on the real hpr4775 pattern: high correlation from the
# start (theme bed, unmasked), a long masked stretch while TTS
# speech dominates (correlation drops but the theme is still
# legitimately playing underneath), then a genuine quiet/
# instrumental-only tail where correlation climbs back up before
# the host's own content begins. The old leading-run algorithm
# would have stopped at the first dip; this must bridge it.
corrs = [0.99] * 24 + [0.2] * 80 + [0.95] * 20 + [0.1] * 20
win_sec = 0.25
result = last_high_correlation_run_seconds(corrs, win_sec)
self.assertAlmostEqual(result, (24 + 80 + 20) * win_sec, places=4)
def test_pure_leading_run_with_no_gap_still_works(self):
corrs = [0.99] * 60 + [0.1] * 40
result = last_high_correlation_run_seconds(corrs, 0.25)
self.assertAlmostEqual(result, 60 * 0.25, places=4)
def test_no_leading_anchor_returns_zero(self):
# The reference isn't correlated from the start at all -- a later
# coincidental high-correlation island must not be trusted as a
# real boundary, since there's no confirmed anchor to bridge from.
corrs = [0.1] * 50 + [0.95] * 10 + [0.1] * 10
result = last_high_correlation_run_seconds(corrs, 0.25)
self.assertEqual(result, 0.0)
def test_single_window_noise_spike_is_ignored(self):
# A lone high-correlation window inside an otherwise-uncorrelated
# stretch shouldn't extend the boundary -- only runs of at least
# min_run_windows count.
corrs = [0.99] * 10 + [0.1] * 30 + [0.97] + [0.1] * 10
result = last_high_correlation_run_seconds(corrs, 0.25)
self.assertAlmostEqual(result, 10 * 0.25, places=4)
if __name__ == "__main__":
unittest.main()