Fix opening filler with bare Instrumental and dual-gate QC.
selftest / selftest (push) Canceled after 0s

ACE-Step no-vocal condition is the exact [Instrumental] marker; colon prose in the lyrics channel induced wordless filler. Recheck now ASRs full vocals and pairs filler ratio with demucs energy.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
swl
2026-10-10 14:29:02 +08:00
co-authored by Cursor
parent 6d79ab93a2
commit 8be5782006
6 changed files with 162 additions and 252 deletions
+1 -1
View File
@@ -117,7 +117,7 @@ def assemble(args, cfg, st, lyrics_raw, title):
st.get("intro_detail") or st.get("instruments"),
bars=ibars)
else:
# 已有 [intro] 头:对齐小节数 + 前奏括号 rhythm 过滤。
# 已有 [intro] 头:收成官方裸 [Instrumental](小节数只在 tags)。
lyrics = ironlaw.force_intro_bars(lyrics, ibars)
lyrics = ironlaw.filter_intro_bracket(lyrics)
lyrics = "\n".join(
+51 -68
View File
@@ -138,7 +138,9 @@ def _t_ironlaw():
lyrics = ironlaw.build_lyrics(
[("[verse]", "第一句歌词\n第二句歌词")], "kick and snare", bars=8)
assert lyrics.lstrip().startswith("[intro]"), "未强制器乐前奏"
assert "[Instrumental: 8 bar" in lyrics, "前奏未写明 8 小节"
assert re.search(r"^\[Instrumental\]\s*$", lyrics, re.M), \
"前奏必须是裸 [Instrumental],不能带冒号散文"
assert "[Instrumental:" not in lyrics, "前奏残留 [Instrumental: ...]"
assert not ironlaw.check_rules(lyrics, tags), ironlaw.check_rules(lyrics, tags)
# 缺前奏必须被拦
no_intro = "[verse]\n没有前奏直接唱"
@@ -191,12 +193,12 @@ def _t_sparse_profile():
"sparse 放行区间 %d~%d 与实测 %d 音节不一致" % (lo, hi, syl)
# 过短仍要拦:sparse 不是"无限制放宽"
thin = ("[intro]\n[Instrumental: 8 bar]\n[verse]\nmmm\nooh\n")
thin = ("[intro]\n[Instrumental]\n[verse]\nmmm\nooh\n")
assert any("过少" in e for e in meter.check(thin, dur, bpm, sparse)[0]), \
"sparse 档放过短歌词"
# 过多仍要拦:写成满首人声时 sparse 也必须报错,否则会退回旧病
fat = ("[intro]\n[Instrumental: 8 bar]\n[verse]\n"
fat = ("[intro]\n[Instrumental]\n[verse]\n"
+ "\n".join(["mmm mmm low and slow"] * 90))
assert any("过多" in e for e in meter.check(fat, dur, bpm, sparse)[0]), \
"sparse 档放过多歌词,音乐主导会被稀释"
@@ -419,22 +421,11 @@ def _t_groove_first_beat():
def _t_intro_instrument_filter():
"""前奏括号必须过滤掉 pad/riser/reverb/lead,只留节奏组。
"""前奏:歌词侧裸 [Instrumental];配器节奏过滤只服务 tags/历史工具。
这是本轮真正起作用的改动。上一轮只在标签里堆禁令,GroovePure 实测
无效(开头 sub 仍 -66.0 dBFS,中段 -55.8 dBFS),因为真正的病灶是
歌词体那一行:
[Instrumental: 8 bar instrumental intro hook, huge supersaw lead,
sidechain compressed pad, punchy electro kick, crisp clap,
trap snare, white noise riser, big reverb, melodic drop, ...]
空场信号(lead/pad/riser/reverb)4 个,节奏组(kick/clap/snare)3 个,
3:4 输了,模型就把前奏铺成 ambient。
所以这里守三件事:
1) 过滤后必须还留着 kick(不能把节奏组也误删成空括号);
2) pad / riser / reverb / lead 一个都不能留在前奏括号里;
3) 标签侧的 instruments 段保持原样——drop 段本来就要 pad 和 lead,
只过滤前奏括号,不能把整首歌的配器改掉。
官方 ACE-Step 无人声条件是精确字符串 [Instrumental]。往括号塞配器
会进歌词通道诱发衬词。intro_rhythm_only 仍保留给 tags / 旧资产,
但 assemble 不再把过滤结果写进歌词。
"""
src = ("huge supersaw lead, sidechain compressed pad, "
"punchy electro kick, crisp clap, trap snare, "
@@ -442,52 +433,43 @@ def _t_intro_instrument_filter():
got = ironlaw.intro_rhythm_only(src)
low = got.lower()
assert "kick" in low, "前奏括号把底鼓也过滤掉了: %r" % got
assert "kick" in low, "节奏过滤把底鼓也删了: %r" % got
for bad in ("pad", "riser", "reverb", "supersaw", "lead", "noise"):
assert bad not in low, "前奏括号残留空场信号 %s: %r" % (bad, got)
assert bad not in low, "节奏过滤残留空场信号 %s: %r" % (bad, got)
assert "bass" in low, "节奏过滤没有 bass:%r" % got
# 加法兜底:kick 和 bass 两条必须显式出现在前奏括号里。
# 上一轮漏了这一条,风格字典里一条贝斯都没写,过滤后只剩三条中高频,
# 结果两版新音频开头 sub 比全曲峰值低 53.5 / 31.2 dB。
assert "bass" in low, "前奏括号没有 bass:%r" % got
# 空 / 中文 / 全被过滤掉 -> 必须回落到默认节奏组,不能留空括号
for probe in ("", None, "ambient pad, huge supersaw lead, big reverb"):
fb = ironlaw.intro_rhythm_only(probe)
assert "kick" in fb.lower(), "前奏括号回落失败 %r -> %r" % (probe, fb)
assert "bass" in fb.lower(), "回落值缺 bass %r -> %r" % (probe, fb)
for bad in ("pad", "riser", "reverb", "supersaw", "lead"):
assert bad not in fb.lower(), "回落值带空场信号 %s: %r" % (bad, fb)
assert "kick" in fb.lower() and "bass" in fb.lower(), \
"节奏过滤回落失败 %r -> %r" % (probe, fb)
# 回落值不能是 DEFAULT_INSTRUMENT:它含 "chord stabs",而 "chord stab"
# 就在 _RHYTHM_DROP 里,回落会把刚剔除掉的空场信号塞回前奏。
assert "stab" not in ironlaw.DEFAULT_INTRO_RHYTHM.lower(), \
"前奏回落值含 chord stab,与过滤规则自相矛盾"
# 标签侧不受影响:instruments 段必须还是原样
tags = ironlaw.build_tags(style="s", instruments=src, mood="m",
instrumental=True)
assert "white noise riser" in tags, \
"过滤逻辑泄漏到了标签侧,drop 段会丢掉 riser"
assert "sidechain compressed pad" in tags, \
"过滤逻辑泄漏到了标签侧,drop 段会丢掉 pad"
assert "white noise riser" in tags, "过滤泄漏到标签侧"
assert "sidechain compressed pad" in tags, "过滤泄漏到标签侧"
# 歌词体里前奏那一行必须既没有 riser/pad,又必须有 bass
lyr = ironlaw.build_inst_lyrics(src, bars=8)
head = lyr.split("\n")[1].lower()
assert "riser" not in head and "pad" not in head, \
"纯器乐歌词体前奏行仍带空场信号: %s" % head
assert "bass" in head, "纯器乐歌词体前奏行缺 bass: %s" % head
lyr2 = ironlaw.build_chill_lyrics([("[verse]", "ooh")], src, bars=8,
total_bars=90)
head2 = lyr2.split("\n")[1].lower()
assert "riser" not in head2 and "pad" not in head2, \
"吟唱歌词体前奏行仍带空场信号: %s" % head2
assert "bass" in head2, "吟唱歌词体前奏行缺 bass: %s" % head2
# 歌词侧:裸标记;旧冒号写法必须被 filter 收成裸
for lyr in (ironlaw.build_inst_lyrics(src, bars=8),
ironlaw.build_chill_lyrics(
[("[verse]", "ooh")], src, bars=8, total_bars=90)):
head = [l.strip() for l in lyr.split("\n") if l.strip()]
assert head[0].lower() == "[intro]"
assert head[1] == "[Instrumental]", \
"歌词前奏不是裸 [Instrumental]: %s" % head[1]
assert ":" not in head[1]
fat = ("[intro]\n[Instrumental: 8 bar hook, pad, riser]\n"
"[instrumental]\n[verse]\nx\n")
fixed = ironlaw.filter_intro_bracket(fat)
assert "[Instrumental: " not in fixed
assert fixed.count("[Instrumental]") == 1
return [("前奏配器过滤",
"前奏括号剔除 pad/riser/reverb/lead,并强制补齐 kick+bass,"
"标签侧配器保持原样")]
"歌词侧裸 [Instrumental];intro_rhythm_only 仍剔 pad/riser;"
"旧冒号写法被 filter 收成裸")]
def _t_chill_no_conflict():
@@ -562,25 +544,26 @@ def _t_chill_integration():
# 再留一遍会让人声提前回来
assert head.count("[inst]") <= n_inst, "chill 结构体器乐占位异常重复"
# 前奏小节数:标签与歌词头必须一致
mh = re.search(r"\[instrumental:[^\]]*?(\d+)\s*bar", lyrics2, re.I)
# 小节数只在 tags;歌词侧必须是裸 [Instrumental]
assert re.search(r"^\[Instrumental\]\s*$", lyrics2, re.M), \
"chill 歌词前奏不是裸 [Instrumental]"
assert "[Instrumental:" not in lyrics2
mt = re.search(r"(\d+)\s*bar instrumental intro", tags2.lower())
assert mh and mt and mh.group(1) == mt.group(1), \
"chill 标签 %s 小节与歌词头 %s 小节不一致" % (
mt.group(1) if mt else "?", mh.group(1) if mh else "?")
# 反向闸门:两者不一致必须被拦下
short_tags = tags2.replace("%s bar instrumental intro" % mt.group(1),
"16 bar instrumental intro", 1)
errs = ironlaw.check_rules(lyrics2, short_tags)
assert any("必须一致" in e for e in errs), \
"tags 与歌词头前奏小节数不一致未被拦截"
assert mt and 4 <= int(mt.group(1)) <= 32, \
"chill tags 缺合法前奏小节数"
# 前奏小节数缺失也必须被拦
errs = ironlaw.check_rules(lyrics2, re.sub(r"\d+\s*bar instrumental",
"instrumental", tags2))
assert any("小节数" in e for e in errs), "tags 未写前奏小节数未被拦截"
# 旧冒号散文必须被拦
bad_head = lyrics2.replace("[Instrumental]",
"[Instrumental: 8 bar pad and riser]", 1)
errs = ironlaw.check_rules(bad_head, tags2)
assert any("裸 [Instrumental]" in e for e in errs), \
"冒号前奏散文未被拦截"
return [("吟唱结构占位",
"人声段之前留了 %d 组纯器乐段(只管密度,不管起点),"
"标签侧完全不讨论人声进场时机" % n_inst)]
@@ -589,7 +572,7 @@ def _t_chill_integration():
def _t_lexicon():
lex = loader.lexicon()
assert "长板" in (lex["banned"].get("longboard") or []), "禁用词表缺 长板"
good = "[intro]\n[Instrumental: 8 bar]\n[verse]\n我在街头双翘上起跳"
good = "[intro]\n[Instrumental]\n[verse]\n我在街头双翘上起跳"
assert not ironlaw.check_lexicon(good, lex), ironlaw.check_lexicon(good, lex)
bad = "[intro]\n[verse]\n我戴着头盔冲下坡"
errs = ironlaw.check_lexicon(bad, lex)
@@ -604,11 +587,11 @@ def _t_density():
assert cap["hi"] > cap["lo"] > 0, "容量区间异常"
# 过短必须被拦
too_thin = "[intro]\n[Instrumental: 8 bar]\n[verse]\n太短了"
too_thin = "[intro]\n[Instrumental]\n[verse]\n太短了"
assert meter.check(too_thin, st["duration"], st["bpm"], cfg)[0], "过短未拦"
# 过多必须被拦
too_fat = "[intro]\n[Instrumental: 8 bar]\n[verse]\n" + \
too_fat = "[intro]\n[Instrumental]\n[verse]\n" + \
"\n".join(["这一句有十二个字刚刚好"] * 300)
assert meter.check(too_fat, st["duration"], st["bpm"], cfg)[0], "过多未拦"
@@ -617,7 +600,7 @@ def _t_density():
line = "这一句刚好十二个字的样子" # 12 音节
nlines = max(1, int(mid / 12))
body = "\n".join([line] * nlines)
ok = "[intro]\n[Instrumental: 8 bar]\n[verse]\n" + body
ok = "[intro]\n[Instrumental]\n[verse]\n" + body
errs, stt = meter.check(ok, st["duration"], st["bpm"], cfg)
assert not errs, "按 target 写的合规歌词被误杀: %s" % errs
return [("时长对应硬校验",
@@ -756,7 +739,7 @@ def _t_window():
line = "这一句刚好十二个字的样子"
def build(n):
return ("[intro]\n[Instrumental: 8 bar]\n[verse]\n"
return ("[intro]\n[Instrumental]\n[verse]\n"
+ "\n".join([line] * n))
n_mid = max(1, int(mid / 12))