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
+16 -11
View File
@@ -23,19 +23,23 @@ def main(job):
bars = int(job.get("intro_bars") or 8)
lang = job.get("lang") or "zh"
intro_sec = bars * 60.0 / bpm * 4.0
# 多取 0.5s 余量
take = min(intro_sec + 0.5, 60.0)
y, sr, ch = q._load_audio(path)
n = int(take * sr)
head = y[:max(1, n)]
a = q.asr(head, sr, lang)
# 与全量 qc 同口径:对整段 vocals 做 ASR,再按时间窗统计。
# 只切前奏短窗喂 Whisper 会严重少计衬词(实测只吐 1 个「嗯」却 PASS)。
v, vsr = q.vocals_of(path, job.get("cache_dir"))
a = q.asr(v, vsr, lang)
segs = a.get("segments") or []
# 时间戳相对 clip 起点,整段都在前奏窗内
stats = q.intro_filler_stats(segs, intro_sec + 1.0, lang)
stats = q.intro_filler_stats(
segs, intro_sec, lang, voc=v, vsr=vsr)
hard = q.HARD
fail = (stats["intro_filler_n"] >= hard["intro_filler_min_chars"]
and stats["intro_filler_ratio"] >= hard["intro_filler_ratio"])
ivr = float(stats.get("intro_voc_rms") or 0)
ifn = int(stats["intro_filler_n"])
ifr = float(stats["intro_filler_ratio"])
# 有人声能量时降低字数门槛:短窗幻觉「单个嗯」也要拦
eff_min = hard["intro_filler_min_chars"]
if ivr >= hard["intro_voc_rms_min"]:
eff_min = min(eff_min, 2)
fail = (ifn >= eff_min and ifr >= hard["intro_filler_ratio"]
and ivr >= hard["intro_voc_rms_min"])
# 前 2s 真唱字(content words)也算开曲违规
early = 0
for s in segs:
@@ -50,6 +54,7 @@ def main(job):
intro_sec=round(intro_sec, 2),
early_content=early,
asr_head=(a.get("text") or "")[:160],
eff_min=eff_min,
**stats,
)
+34 -11
View File
@@ -86,6 +86,9 @@ HARD = dict(
# 阈值刻意偏严:用户要求「不能出开曲衬词」;Whisper 偶发单字可被 min 挡住。
intro_filler_min_chars=6, # 前奏窗内至少累计这么多衬词字/词才判
intro_filler_ratio=0.55, # 衬词占前奏窗 ASR 单元比例
# Whisper 常在纯器乐上幻觉「嗯/啊」;必须再看 demucs vocals 能量。
# 前奏窗 vocals RMS 低于此阈值 → 视为幻觉,不因 ASR 衬词 FAIL。
intro_voc_rms_min=0.008,
# ---- 歌词感知分类阈值(见文件头坑位 6/7)----
# lyric_hit >= 0.34 -> 真唱;instr_hit >= 0.40 -> 生成指令泄漏。
# 这两个数是在 8 首已认可老歌 + 本批 8 首上交叉验证出来的:
@@ -706,11 +709,14 @@ def filler_ratio(asr_text, lang="zh"):
return n / float(len(chars))
def intro_filler_stats(segs, intro_sec, lang="zh"):
def intro_filler_stats(segs, intro_sec, lang="zh", voc=None, vsr=None):
"""前奏时间窗内的衬词统计(硬门用)。
vocal_onset / head_chars 只看「真唱实义词」,会故意丢掉 嗯/呀/ooh,
导致开曲衬词满天飞仍 PASS。这里单独扫前奏窗 ASR。
若传入 demucs vocals(voc/vsr),同时算前奏窗 RMS:Whisper 在纯器乐上
常幻觉衬词,能量门用来区分「真有无词人声」与「ASR 幻觉」。
"""
total = 0
fill = 0
@@ -732,9 +738,15 @@ def intro_filler_stats(segs, intro_sec, lang="zh"):
"".join(f)[:24] if not _is_en(lang)
else " ".join(f)[:24]))
ratio = (fill / float(total)) if total else 0.0
intro_voc_rms = 0.0
if voc is not None and vsr:
n = int(min(float(intro_sec), len(voc) / float(vsr)) * float(vsr))
if n > 0:
intro_voc_rms = _rms(voc[:n])
return dict(intro_filler_n=fill, intro_asr_n=total,
intro_filler_ratio=round(ratio, 4),
intro_filler_samples=samples[:4])
intro_filler_samples=samples[:4],
intro_voc_rms=round(intro_voc_rms, 6))
def coverage(asr_text, lyrics, lang="zh"):
@@ -933,9 +945,10 @@ def measure(path, expect=None, cache_dir=None, with_asr=True):
if expect.get("bpm") and expect.get("intro_bars"):
intro_sec = (float(expect["intro_bars"]) * 60.0
/ float(expect["bpm"]) * 4.0)
ifest = intro_filler_stats(segs, intro_sec, lang) if with_asr else \
ifest = intro_filler_stats(segs, intro_sec, lang, voc=v, vsr=vsr) \
if with_asr else \
dict(intro_filler_n=0, intro_asr_n=0, intro_filler_ratio=0.0,
intro_filler_samples=[])
intro_filler_samples=[], intro_voc_rms=0.0)
# ---- 风格分类:只看混音轨,不掺 demucs(见 qc_style 文件头坑位 4)
svec, sraw = qc_style.features(path, expect_bpm=expect.get("bpm"))
@@ -967,6 +980,7 @@ def measure(path, expect=None, cache_dir=None, with_asr=True):
intro_asr_n=ifest["intro_asr_n"],
intro_filler_ratio=ifest["intro_filler_ratio"],
intro_filler_samples=ifest["intro_filler_samples"],
intro_voc_rms=ifest.get("intro_voc_rms", 0.0),
asr_head=txt[:120],
voice_self=round(float(e_self), 4),
flatness=round(sraw.get("flatness", 0.0), 4),
@@ -1069,18 +1083,27 @@ def judge(rows, baseline):
why.append("前 2s 就已唱出 %d 个%s,疑似纯人声/spoken 开头"
% (r["head_chars_0_2"], unit))
# ---- 铁律一-b:禁止开曲衬词(嗯/呀/ooh…)
# vocal_onset 只看真唱,会漏掉衬词;此处扫前奏窗 ASR。
# ASR 衬词 + demucs vocals 能量双门:纯器乐静音上的幻觉不 FAIL。
# demucs 常有 ~0.09 器乐泄漏,能量门挡的是近零 RMS;有能量时
# 字数门槛降到 2,避免「满屏嗯」却因短窗少计而漏网。
ifn = int(r.get("intro_filler_n") or 0)
ifr = float(r.get("intro_filler_ratio") or 0.0)
if ifn >= HARD["intro_filler_min_chars"] and \
ifr >= HARD["intro_filler_ratio"]:
ivr = float(r.get("intro_voc_rms") or 0.0)
eff_min = HARD["intro_filler_min_chars"]
if ivr >= HARD["intro_voc_rms_min"]:
eff_min = min(eff_min, 2)
if ifn >= eff_min and ifr >= HARD["intro_filler_ratio"] and \
ivr >= HARD["intro_voc_rms_min"]:
samp = r.get("intro_filler_samples") or []
tip = ("; ".join("%ss:%s" % (a, b) for a, b in samp[:2])
if samp else "")
why.append("开曲衬词 %d 单元占比 %.0f%%(≥%d 且 ≥%.0f%%)%s"
% (ifn, ifr * 100, HARD["intro_filler_min_chars"],
HARD["intro_filler_ratio"] * 100,
(":" + tip) if tip else ""))
why.append(
"开曲衬词 %d 单元占比 %.0f%% 且前奏人声能量 %.4f"
"(≥%d 且 ≥%.0f%% 且 RMS≥%.3f)%s"
% (ifn, ifr * 100, ivr, eff_min,
HARD["intro_filler_ratio"] * 100,
HARD["intro_voc_rms_min"],
(":" + tip) if tip else ""))
# ---- 铁律二:8 小节器乐前奏(完整 8 小节,不是 55%)
# 原来判的是 vocal_onset < want * 0.55,等于"第 4~5 小节进 sang
# 也算过",把铁律降级成建议。现在按完整 8 小节判,只留