refactor(scanner): use shared verification core; stop false-flagging cross-script
The library AcoustID scan now calls audio_verification.evaluate() (alias-aware artist match + cross-script SKIP) instead of its own non-ASCII-stripping _normalize and threshold logic, so it no longer false-flags correct anime-OST / kanji tracks. Duration-collision guard kept as a scanner pre-check on the top recording. evaluate() is now purely a title/artist/version/cross-script decision. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
dev
parent
967ad2a026
commit
b981230d07
3 changed files with 95 additions and 124 deletions
|
|
@ -159,20 +159,18 @@ def _find_best_title_artist_match(recordings, expected_title, expected_artist,
|
|||
|
||||
def evaluate(expected_title: str, expected_artist: str,
|
||||
recordings: List[dict], *, fingerprint_score: float,
|
||||
file_duration_s: Optional[float] = None,
|
||||
aliases_provider: Optional[Any] = None) -> Outcome:
|
||||
"""Decide PASS / SKIP / FAIL for a fingerprinted file against expected
|
||||
title/artist. Pure: no I/O. Shared by import verification and library scan.
|
||||
|
||||
``aliases_provider``: iterable or callable of expected-artist aliases
|
||||
(kanji/cyrillic/etc) used to bridge cross-script comparisons.
|
||||
``file_duration_s``: when provided, a strong duration mismatch downgrades a
|
||||
would-be FAIL to SKIP (fingerprint hash collision guard, used by the scan).
|
||||
|
||||
Note: fingerprint-collision duration checks are the caller's responsibility
|
||||
(the library scan pre-checks the top recording's length before calling this)
|
||||
so the decision here stays purely about title/artist/version identity.
|
||||
"""
|
||||
from core.matching.script_compat import is_cross_script_mismatch
|
||||
from core.matching.acoustid_candidates import (
|
||||
duration_mismatches_strongly, find_matching_recording,
|
||||
)
|
||||
from core.matching.version_mismatch import is_acceptable_version_mismatch
|
||||
|
||||
best_rec, title_sim, artist_sim = _find_best_title_artist_match(
|
||||
|
|
@ -212,9 +210,6 @@ def evaluate(expected_title: str, expected_artist: str,
|
|||
) >= ARTIST_MATCH_THRESHOLD:
|
||||
return out(Decision.PASS, "Expected artist found in AcoustID results")
|
||||
if artist_sim < CLEAR_MISMATCH_THRESHOLD:
|
||||
if file_duration_s and duration_mismatches_strongly(
|
||||
file_duration_s, best_rec.get('duration') or best_rec.get('length')):
|
||||
return out(Decision.SKIP, "Duration mismatch (fingerprint collision)")
|
||||
return out(Decision.FAIL,
|
||||
f"Audio mismatch: '{matched_title}' by '{matched_artist}' "
|
||||
f"— expected artist not found")
|
||||
|
|
@ -255,10 +250,6 @@ def evaluate(expected_title: str, expected_artist: str,
|
|||
or cross_script_artist_skip):
|
||||
return out(Decision.SKIP, "Likely same song in different language/script")
|
||||
|
||||
if file_duration_s and duration_mismatches_strongly(
|
||||
file_duration_s, best_rec.get('duration') or best_rec.get('length')):
|
||||
return out(Decision.SKIP, "Duration mismatch (fingerprint collision)")
|
||||
|
||||
return out(Decision.FAIL,
|
||||
f"Audio mismatch: file identified as '{matched_title}' by "
|
||||
f"'{matched_artist}', expected '{expected_title}' by '{expected_artist}'")
|
||||
|
|
|
|||
|
|
@ -15,6 +15,9 @@ from typing import Optional
|
|||
from core.repair_jobs import register_job
|
||||
from core.repair_jobs.base import JobContext, JobResult, RepairJob
|
||||
from utils.logging_config import get_logger
|
||||
from core.matching.audio_verification import evaluate, Decision
|
||||
from core.matching.acoustid_candidates import duration_mismatches_strongly
|
||||
from core.acoustid_verification import _resolve_expected_artist_aliases
|
||||
|
||||
logger = get_logger("repair_job.acoustid")
|
||||
|
||||
|
|
@ -220,112 +223,60 @@ class AcoustIDScannerJob(RepairJob):
|
|||
or expected['artist']
|
||||
)
|
||||
|
||||
# Normalize and compare
|
||||
norm_expected_title = _normalize(expected['title'])
|
||||
norm_aid_title = _normalize(aid_title)
|
||||
norm_expected_artist = _normalize(expected_artist)
|
||||
norm_aid_artist = _normalize(aid_artist)
|
||||
|
||||
title_sim = SequenceMatcher(None, norm_expected_title, norm_aid_title).ratio()
|
||||
# Issue (Foxxify Discord report): AcoustID returns the FULL artist
|
||||
# credit (e.g. `Okayracer, aldrch & poptropicaslutz!`) while the
|
||||
# library DB carries only the primary artist (`Okayracer`). Raw
|
||||
# similarity scores ~43% — well below threshold — so multi-artist
|
||||
# tracks get flagged as Wrong Song even though the primary IS in
|
||||
# the credit. Route through the shared `artist_names_match` helper
|
||||
# which splits the credit on common separators (comma, ampersand,
|
||||
# feat./ft./with/vs., etc.) and checks each token. Primary-in-
|
||||
# credit cases now resolve at 100% match instead of 43%.
|
||||
#
|
||||
# Pass RAW artist strings (not pre-normalised) so the splitter
|
||||
# can recognise the separators. The helper applies its own
|
||||
# case + whitespace normalisation internally per token.
|
||||
if norm_expected_artist:
|
||||
from core.matching.artist_aliases import artist_names_match
|
||||
|
||||
_, artist_sim = artist_names_match(
|
||||
expected_artist,
|
||||
aid_artist,
|
||||
threshold=artist_threshold,
|
||||
)
|
||||
else:
|
||||
artist_sim = 1.0
|
||||
|
||||
if title_sim >= title_threshold and artist_sim >= artist_threshold:
|
||||
return
|
||||
|
||||
# Issue #587 (Foxxify) — top recording's metadata mismatched, but
|
||||
# AcoustID often returns multiple recordings per fingerprint
|
||||
# (sample collisions, multi-MB-record cases). Check ALL of them
|
||||
# before flagging — if any candidate's metadata matches expected
|
||||
# title + artist, the file IS the right song and AcoustID's top
|
||||
# match was just a wrong-credited recording.
|
||||
from core.matching.acoustid_candidates import (
|
||||
duration_mismatches_strongly,
|
||||
find_matching_recording,
|
||||
)
|
||||
from core.matching.artist_aliases import artist_names_match
|
||||
|
||||
def _scanner_title_sim(a, b):
|
||||
return SequenceMatcher(None, _normalize(a), _normalize(b)).ratio()
|
||||
|
||||
def _scanner_artist_sim(expected_a, actual_a):
|
||||
_, score = artist_names_match(expected_a, actual_a, threshold=artist_threshold)
|
||||
return score
|
||||
|
||||
candidate_match, _, _ = find_matching_recording(
|
||||
fp_result.get('recordings') or [],
|
||||
expected['title'],
|
||||
expected_artist,
|
||||
title_threshold=title_threshold,
|
||||
artist_threshold=artist_threshold,
|
||||
similarity=_scanner_title_sim,
|
||||
artist_similarity=_scanner_artist_sim,
|
||||
)
|
||||
if candidate_match is not None:
|
||||
# A lower-ranked candidate matched — file IS the right song.
|
||||
# No finding.
|
||||
# Fingerprint-collision guard: when the TOP recording's length is wildly
|
||||
# different from the file, the fingerprint hit is a hash collision (the
|
||||
# 17-min mashup → 5-min track case), not a real match — skip BEFORE any
|
||||
# title/artist/version analysis so it can't surface as a false finding.
|
||||
try:
|
||||
file_duration_s = (expected.get('duration_ms') or 0) / 1000.0
|
||||
except Exception:
|
||||
file_duration_s = 0.0
|
||||
cand_duration_s = best_recording.get('duration') or best_recording.get('length')
|
||||
if file_duration_s and duration_mismatches_strongly(file_duration_s, cand_duration_s):
|
||||
if context.report_progress:
|
||||
context.report_progress(
|
||||
log_line=(
|
||||
f'Resolved (lower-ranked candidate match): {fname} — '
|
||||
f'expected "{expected["title"]}" matched candidate '
|
||||
f'"{candidate_match.get("title")}" by '
|
||||
f'"{candidate_match.get("artist")}"'
|
||||
),
|
||||
log_line=(f'Skipped (duration mismatch suggests fingerprint '
|
||||
f'collision): {fname}'),
|
||||
log_type='skip')
|
||||
return
|
||||
|
||||
# Decision via the shared verification core — identical logic to import-
|
||||
# time verification (alias-aware artist match + cross-script SKIP), so the
|
||||
# scan no longer false-flags correct cross-script tracks. Only a FAIL
|
||||
# produces a "Wrong download" finding.
|
||||
_alias_cache = {}
|
||||
|
||||
def _aliases():
|
||||
if 'v' not in _alias_cache:
|
||||
try:
|
||||
_alias_cache['v'] = _resolve_expected_artist_aliases(expected_artist)
|
||||
except Exception:
|
||||
_alias_cache['v'] = []
|
||||
return _alias_cache['v']
|
||||
|
||||
outcome = evaluate(
|
||||
expected['title'], expected_artist, fp_result['recordings'],
|
||||
fingerprint_score=best_score,
|
||||
aliases_provider=_aliases,
|
||||
)
|
||||
|
||||
if outcome.decision != Decision.FAIL:
|
||||
if context.report_progress:
|
||||
context.report_progress(
|
||||
log_line=f'OK ({outcome.decision.value}): {fname} — {outcome.reason}',
|
||||
log_type='ok',
|
||||
)
|
||||
return
|
||||
|
||||
# Issue #587 (Foxxify "17min mashup → 5min track") — duration
|
||||
# guard against fingerprint hash collisions. When the file's
|
||||
# actual duration differs from AcoustID's matched recording by
|
||||
# more than max(60s, 35%), the fingerprint is almost certainly
|
||||
# a sample/intro collision, not a real recording match. Don't
|
||||
# produce a confident "Wrong Song" finding.
|
||||
try:
|
||||
file_duration_s = (expected.get('duration_ms') or 0) / 1000.0
|
||||
except Exception:
|
||||
file_duration_s = 0
|
||||
candidate_duration_s = best_recording.get('duration')
|
||||
if candidate_duration_s is None and best_recording.get('length'):
|
||||
candidate_duration_s = best_recording.get('length')
|
||||
if duration_mismatches_strongly(file_duration_s, candidate_duration_s):
|
||||
if context.report_progress:
|
||||
context.report_progress(
|
||||
log_line=(
|
||||
f'Skipped (duration mismatch suggests fingerprint collision): '
|
||||
f'{fname} — expected {file_duration_s:.0f}s, AcoustID '
|
||||
f'candidate {candidate_duration_s:.0f}s'
|
||||
),
|
||||
log_type='skip',
|
||||
)
|
||||
return
|
||||
title_sim = outcome.title_sim
|
||||
artist_sim = outcome.artist_sim
|
||||
matched_title = outcome.matched_title or aid_title
|
||||
matched_artist = outcome.matched_artist or aid_artist
|
||||
|
||||
# Mismatch detected
|
||||
# Mismatch (FAIL) — create finding.
|
||||
if context.report_progress:
|
||||
context.report_progress(
|
||||
log_line=f'Mismatch: {fname} — expected "{expected["title"]}", got "{aid_title}"',
|
||||
log_line=f'Mismatch: {fname} — expected "{expected["title"]}", got "{matched_title}"',
|
||||
log_type='error'
|
||||
)
|
||||
if context.create_finding:
|
||||
|
|
@ -337,18 +288,18 @@ class AcoustIDScannerJob(RepairJob):
|
|||
entity_type='track',
|
||||
entity_id=str(track_id),
|
||||
file_path=fpath,
|
||||
title=f'Wrong download: "{expected["title"]}" is actually "{aid_title}"',
|
||||
title=f'Wrong download: "{expected["title"]}" is actually "{matched_title}"',
|
||||
description=(
|
||||
f'Expected "{expected["title"]}" by {expected_artist}, '
|
||||
f'but audio fingerprint matches "{aid_title}" by {aid_artist} '
|
||||
f'but audio fingerprint matches "{matched_title}" by {matched_artist} '
|
||||
f'(fingerprint: {best_score:.0%}, title match: {title_sim:.0%}, '
|
||||
f'artist match: {artist_sim:.0%})'
|
||||
),
|
||||
details={
|
||||
'expected_title': expected['title'],
|
||||
'expected_artist': expected_artist,
|
||||
'acoustid_title': aid_title,
|
||||
'acoustid_artist': aid_artist,
|
||||
'acoustid_title': matched_title,
|
||||
'acoustid_artist': matched_artist,
|
||||
'fingerprint_score': round(best_score, 3),
|
||||
'title_similarity': round(title_sim, 3),
|
||||
'artist_similarity': round(artist_sim, 3),
|
||||
|
|
@ -480,11 +431,3 @@ class AcoustIDScannerJob(RepairJob):
|
|||
finally:
|
||||
if conn:
|
||||
conn.close()
|
||||
|
||||
|
||||
def _normalize(text: str) -> str:
|
||||
t = text.lower()
|
||||
t = re.sub(r'\(.*?\)', '', t)
|
||||
t = re.sub(r'\[.*?\]', '', t)
|
||||
t = re.sub(r'[^a-z0-9 ]', '', t)
|
||||
return t.strip()
|
||||
|
|
|
|||
|
|
@ -199,7 +199,11 @@ def test_scanner_still_flags_genuine_artist_mismatch():
|
|||
'best_score': 0.99,
|
||||
'recordings': [{
|
||||
'title': 'Some Track',
|
||||
'artist': 'Different Band, Other Person & Random Featuring',
|
||||
# Clearly-different multi-value credit (artist sim < 0.30). The
|
||||
# unified core gives 0.30-0.60 ("ambiguous") the benefit of the
|
||||
# doubt, so a genuine-mismatch assertion needs an artist that's
|
||||
# unambiguously different.
|
||||
'artist': 'Metallica, Slayer & Anthrax',
|
||||
}],
|
||||
},
|
||||
)
|
||||
|
|
@ -468,7 +472,9 @@ def test_scanner_falls_back_to_db_when_file_tag_missing(monkeypatch):
|
|||
'best_score': 0.99,
|
||||
'recordings': [{
|
||||
'title': 'Some Track',
|
||||
'artist': 'Different Band',
|
||||
# Unambiguously different artist (sim < 0.30) so the unified
|
||||
# core flags it (0.30-0.60 would be treated as ambiguous).
|
||||
'artist': 'Metallica',
|
||||
}],
|
||||
},
|
||||
)
|
||||
|
|
@ -779,3 +785,34 @@ def test_scanner_still_flags_when_duration_matches():
|
|||
)
|
||||
|
||||
assert len(captured_findings) == 1
|
||||
|
||||
|
||||
def test_scanner_does_not_flag_cross_script_when_alias_bridges(monkeypatch):
|
||||
"""Anime-OST track: AcoustID returns the kanji artist with a <Vocal: ...>
|
||||
credit. With the MusicBrainz alias bridging 澤野弘之 ↔ Sawano Hiroyuki, the
|
||||
unified verification core recognises the match, so the library scan must NOT
|
||||
create a false 'Wrong download' finding (it did before, stripping all
|
||||
non-ASCII and never consulting aliases)."""
|
||||
import core.repair_jobs.acoustid_scanner as scanner_mod
|
||||
monkeypatch.setattr(scanner_mod, "_resolve_expected_artist_aliases",
|
||||
lambda name: ["澤野弘之"], raising=False)
|
||||
job = AcoustIDScannerJob()
|
||||
captured = []
|
||||
context = _make_finding_capturing_context(
|
||||
track_row=("7", "Call Your Name", "Sawano Hiroyuki",
|
||||
"/music/cyn.flac", 15, "Attack on Titan OST", None, None),
|
||||
captured=captured,
|
||||
)
|
||||
fake_acoustid = SimpleNamespace(
|
||||
fingerprint_and_lookup=lambda fpath: {
|
||||
'best_score': 0.97,
|
||||
'recordings': [{'title': 'call your name',
|
||||
'artist': '澤野弘之 <Vocal: mpi & CASG>'}],
|
||||
},
|
||||
)
|
||||
result = JobResultStub()
|
||||
job._scan_file('/music/cyn.flac', '7',
|
||||
{'title': 'Call Your Name', 'artist': 'Sawano Hiroyuki'},
|
||||
fake_acoustid, context, result,
|
||||
fp_threshold=0.85, title_threshold=0.85, artist_threshold=0.6)
|
||||
assert captured == [], f"cross-script track false-flagged: {captured}"
|
||||
|
|
|
|||
Loading…
Reference in a new issue