Add songforge as independent repo with path layout.
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
@@ -0,0 +1,416 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""风格特征与风格分类 —— 只吃 librosa,纯数值,不碰 demucs/whisper。
|
||||
|
||||
为什么单独一个文件:
|
||||
songforge 主链路是零重依赖的,风格判定属于"生成后验收"的一部分,
|
||||
必须在 venv 侧跑。把它独立出来,`qc_worker.py` 和
|
||||
`tools/make_style_ref.py` 共用同一套取特征逻辑——否则"建语料"和
|
||||
"验收"各写一份,特征对不上,分类结果就是自欺欺人。
|
||||
|
||||
风格怎么判(重要,不要改成拍脑袋的阈值):
|
||||
这里不做"频谱质心 > X 就是朋克"这种单指标硬判。做法是拿历史里
|
||||
已经按风格命名、并且用户听过没否掉的歌当语料,逐首算同一套特征,
|
||||
按稳健标准化后建三个风格质心(pop / punk / hiphop),用最近质心
|
||||
分类,并做留一验证(LOO)把准确率一起存下来。
|
||||
准确率不够高时,报告里如实写"风格判定不可信",只当参考不当闸门。
|
||||
|
||||
坑位记录:
|
||||
1. 速度必须做八度折叠。实测 `晚风不加班`(设定 86 BPM)会被判成
|
||||
170.5 BPM,正好是 2 倍——librosa 的 onset 包络在半速拍子上会
|
||||
锁到二倍拍。直接比设定值会把这个风格判死。
|
||||
2. 频谱类特征跨度极大(质心 1400~6000 Hz),必须取 log 再标准化,
|
||||
否则质心一项就主导了整段欧氏距离。
|
||||
3. perc_ratio 用 HPSS 的打击成分占比。朋克的军鼓+失真吉他、
|
||||
hiphop 的 kick+hi-hat 都会偏高,流行的人声+弦乐偏低。
|
||||
单看这个不如 BPM 有区分度,但作为辅助项能压掉误判。
|
||||
4. 特征里绝对不能有"人声轨 / 混音轨 能量比"。带标签的历史语料有 30~50 首,
|
||||
每一首都跑 demucs 要花几十分钟,而且 demucs 对不同混音的分离程度
|
||||
本身就不一致,这个比值混进去是纯噪声。风格只看混音轨。
|
||||
"""
|
||||
import os
|
||||
import re
|
||||
|
||||
import numpy as np
|
||||
|
||||
FAMILY = {
|
||||
"pop": "pop",
|
||||
"punk": "punk",
|
||||
"emopunk": "punk",
|
||||
"urbanhiphop": "hiphop",
|
||||
"chillhiphop": "hiphop",
|
||||
"partyrap": "hiphop",
|
||||
"funkhiphop": "hiphop",
|
||||
"discofunk": "hiphop",
|
||||
"jazzhiphop": "hiphop",
|
||||
"guofeng": "hiphop",
|
||||
}
|
||||
|
||||
FAMILIES = ("pop", "punk", "hiphop")
|
||||
|
||||
# 特征顺序即向量顺序,改这里必须重建 style_ref.json,否则老语料错位。
|
||||
FEATURES = (
|
||||
"tempo", # 八度折叠后的实测速度
|
||||
"tempo_drift", # 速度稳定度(越小越稳)
|
||||
"centroid", # log 频谱质心 -> 亮/暗
|
||||
"rolloff", # log 85% rolloff -> 高频能量
|
||||
"hf_ratio", # >6kHz 能量占比 -> 亮泽/嘶叫
|
||||
"lf_ratio", # <150Hz 能量占比 -> 低频厚度/sub
|
||||
"dynamic", # 动态范围 dB
|
||||
"onset_density", # 每秒击打数
|
||||
"perc_ratio", # HPSS 打击成分占比
|
||||
"flatness", # 谱平坦度 -> 噪声/失真成分(朋克嘶叫会抬高)
|
||||
"crest", # 峰值因子,越小越被压得平
|
||||
)
|
||||
|
||||
_LOG_IDX = (FEATURES.index("centroid"), FEATURES.index("rolloff"))
|
||||
|
||||
|
||||
# ------------------------------------------------------------------ 工具
|
||||
|
||||
def fold_bpm(value, ref):
|
||||
"""把 value 折叠到离 ref 最近的八度。
|
||||
|
||||
坑位:librosa.beat 在慢速律动上会锁到 2 倍拍,不折叠会把 86 BPM 的歌
|
||||
判成 170 BPM,然后"实测速度与设定偏差 98%",一个完全正常的歌被判死。
|
||||
"""
|
||||
if value is None or ref is None:
|
||||
return value
|
||||
try:
|
||||
v = float(value)
|
||||
r = float(ref)
|
||||
except (TypeError, ValueError):
|
||||
return value
|
||||
# 坑位:非有限值和非正数会一路穿到 log2 里,log2(inf) 再喂给 round()
|
||||
# 直接抛 OverflowError,把整首 QC 打断。下游只想要"折不动就原样返回"。
|
||||
if not (np.isfinite(v) and np.isfinite(r)) or v <= 0 or r <= 0:
|
||||
return v
|
||||
ratio = r / v
|
||||
if not np.isfinite(ratio) or ratio <= 0:
|
||||
return v
|
||||
return v * (2.0 ** round(np.log2(ratio)))
|
||||
|
||||
|
||||
def fold_track(track, ref):
|
||||
return np.array([fold_bpm(x, ref) for x in track], dtype=np.float32)
|
||||
|
||||
|
||||
def _band_ratio(y, sr, lo, hi):
|
||||
"""指定频带能量占比。用 rfft 直接算,比 mel 滤波器组省事也够稳。"""
|
||||
if y.size < 256:
|
||||
return 0.0
|
||||
n = 1 << int(np.ceil(np.log2(y.size)))
|
||||
sp = np.abs(np.fft.rfft(y, n)) ** 2
|
||||
freqs = np.fft.rfftfreq(n, 1.0 / sr)
|
||||
band = sp[(freqs >= lo) & (freqs < hi)].sum()
|
||||
tot = sp.sum()
|
||||
return float(band / tot) if tot > 0 else 0.0
|
||||
|
||||
|
||||
# ------------------------------------------------------------------ 取特征
|
||||
|
||||
def tempo_of(y, sr, expect_bpm=None):
|
||||
"""返回 (折叠后中位速度, 折叠后速度漂移)。expect_bpm 用来定折叠方向。"""
|
||||
import librosa
|
||||
y16 = y if sr == 16000 else librosa.resample(
|
||||
np.asarray(y, dtype=np.float32), orig_sr=sr, target_sr=16000)
|
||||
sr16 = 16000 if sr != 16000 else sr
|
||||
oe = librosa.onset.onset_strength(y=y16, sr=sr16)
|
||||
# librosa 0.11 起 beat.tempo 全是 keyword-only,位置传参直接 TypeError
|
||||
tr = librosa.beat.tempo(onset_envelope=oe, sr=sr16, hop_length=512,
|
||||
aggregate=None)
|
||||
tr = np.asarray(tr, dtype=np.float32)
|
||||
tr = tr[np.isfinite(tr) & (tr > 0)]
|
||||
if not tr.size:
|
||||
return 0.0, 0.0
|
||||
raw_med = float(np.median(tr))
|
||||
anchor = fold_bpm(raw_med, expect_bpm) if expect_bpm else raw_med
|
||||
ft = fold_track(tr, anchor)
|
||||
ft = ft[(ft > 0) & (ft < 400)]
|
||||
if not ft.size:
|
||||
return 0.0, 0.0
|
||||
med = float(np.median(ft))
|
||||
# 坑位:原来用 (p90 - p10) / median 算漂移,但 librosa 逐帧 tempo 的首尾
|
||||
# 各有一拍宽的边界伪影,180s 的歌能攒出十几秒伪影区,被 p90/p10 一夹
|
||||
# 就放大成 25%+ 的"漂移",而同一首歌按 10s 分块的中位全是设定值
|
||||
# (实测 104.2/104.2,IQR=0)。改用中位绝对偏差,只对整体性的加速/
|
||||
# 减速敏感,对边界伪影和偶发跳拍免疫。
|
||||
drift = _mad_rel(ft, med)
|
||||
return med, drift
|
||||
|
||||
|
||||
def _mad_rel(track, med):
|
||||
"""相对中位绝对偏差。返回 0.0 表示测不出整体漂移。"""
|
||||
if track.size == 0 or med <= 0:
|
||||
return 0.0
|
||||
mad = float(np.median(np.abs(track - med)))
|
||||
return (mad / med) if med > 0 else 0.0
|
||||
|
||||
|
||||
def features(path, expect_bpm=None, cache_dir=None):
|
||||
"""算一首歌的完整风格特征向量。
|
||||
|
||||
坑位:这个函数只允许吃混音轨(path)。不要往里塞 demucs 人声轨——
|
||||
建历史语料时要对几十首老歌跑,塞进去就等于把风格判定绑死在分离质量上,
|
||||
而 demucs 的分离程度本身随混音变化。声线由 qc_worker 单独负责。
|
||||
"""
|
||||
import librosa
|
||||
import soundfile as sf
|
||||
|
||||
x, sr = sf.read(path, always_2d=True)
|
||||
x = np.asarray(x, dtype=np.float32)
|
||||
y = x.mean(axis=1)
|
||||
|
||||
dur = y.size / float(sr)
|
||||
tempo, tempo_drift = tempo_of(y, sr, expect_bpm)
|
||||
|
||||
y16 = librosa.resample(y, orig_sr=sr, target_sr=16000)
|
||||
sr16 = 16000
|
||||
S = np.abs(librosa.stft(y16, n_fft=2048, hop_length=512)) + 1e-12
|
||||
centroid = float(np.mean(librosa.feature.spectral_centroid(S=S, sr=sr16)))
|
||||
rolloff = float(np.mean(librosa.feature.spectral_rolloff(
|
||||
S=S, sr=sr16, roll_percent=0.85)))
|
||||
# 谱平坦度:几何均值/算术均值。噪声化(失真、嘶叫、lo-fi)会把它抬高,
|
||||
# 干净的钢琴/弦乐 pad 会压得很低。比 "高频能量占比" 更能区分失真音色。
|
||||
gm = np.exp(np.mean(np.log(S), axis=0))
|
||||
am = np.mean(S, axis=0)
|
||||
flatness = float(np.mean(gm / am))
|
||||
|
||||
bs = 3 * sr
|
||||
n = max(1, y.size // bs)
|
||||
rmss = np.array([np.sqrt(np.mean(np.square(y[i * bs:(i + 1) * bs])) + 1e-20)
|
||||
for i in range(n)], dtype=np.float32)
|
||||
dyn = float(20 * np.log10(max(rmss.max(), 1e-9) / max(rmss.min(), 1e-9)))
|
||||
|
||||
oe = librosa.onset.onset_strength(y=y16, sr=sr16)
|
||||
hop = 512
|
||||
onset_density = float(len(librosa.onset.onset_detect(
|
||||
onset_envelope=oe, sr=sr16, hop_length=hop)) / max(dur, 1e-6))
|
||||
|
||||
h, p = librosa.effects.hpss(y16)
|
||||
perc_ratio = float(np.mean(np.abs(p)) /
|
||||
max(np.mean(np.abs(h)) + np.mean(np.abs(p)), 1e-9))
|
||||
|
||||
peak = float(np.max(np.abs(y))) if y.size else 0.0
|
||||
rms_all = float(np.sqrt(np.mean(np.square(y)) + 1e-20))
|
||||
crest = float(20 * np.log10(max(peak, 1e-9) / max(rms_all, 1e-9)))
|
||||
|
||||
vec = {
|
||||
"tempo": tempo,
|
||||
"tempo_drift": tempo_drift,
|
||||
"centroid": float(np.log10(max(centroid, 1.0))),
|
||||
"rolloff": float(np.log10(max(rolloff, 1.0))),
|
||||
"hf_ratio": _band_ratio(y16, sr16, 6000, sr16 / 2),
|
||||
"lf_ratio": _band_ratio(y16, sr16, 20, 150),
|
||||
"dynamic": dyn,
|
||||
"onset_density": onset_density,
|
||||
"perc_ratio": perc_ratio,
|
||||
"flatness": flatness,
|
||||
"crest": crest,
|
||||
}
|
||||
return vec, dict(centroid_hz=centroid, rolloff_hz=rolloff,
|
||||
flatness=flatness,
|
||||
onset_density=onset_density, perc_ratio=perc_ratio,
|
||||
crest=crest, peak=peak)
|
||||
|
||||
|
||||
def vector(vec):
|
||||
return np.array([float(vec[k]) for k in FEATURES], dtype=np.float64)
|
||||
|
||||
|
||||
# ------------------------------------------------------------------ 分类器
|
||||
|
||||
def _robust_scale(X):
|
||||
"""中位数 + IQR 标准化。IQR 比 std 抗离群值。"""
|
||||
med = np.median(X, axis=0)
|
||||
q1 = np.percentile(X, 25, axis=0)
|
||||
q3 = np.percentile(X, 75, axis=0)
|
||||
iqr = np.maximum(q3 - q1, 1e-6)
|
||||
return med, iqr
|
||||
|
||||
|
||||
class StyleRef(object):
|
||||
"""风格质心分类器。语料不足或 LOO 太低时 self.trusted=False。"""
|
||||
|
||||
def __init__(self, data):
|
||||
self.note = data.get("note", "")
|
||||
self.loo = data.get("loo", {})
|
||||
self.items = data.get("items", [])
|
||||
self.med = np.array(data["med"], dtype=np.float64)
|
||||
self.iqr = np.array(data["iqr"], dtype=np.float64)
|
||||
self.centroids = {k: np.array(v, dtype=np.float64)
|
||||
for k, v in data["centroids"].items()}
|
||||
self.counts = data.get("counts", {})
|
||||
self.trusted = bool(data.get("trusted", False))
|
||||
|
||||
def _z(self, vec):
|
||||
return (vector(vec) - self.med) / self.iqr
|
||||
|
||||
def classify(self, vec):
|
||||
"""返回 (top_family, conf, dists)。conf 是 softmax 后的置信度。"""
|
||||
z = self._z(vec)
|
||||
dists = {}
|
||||
for fam, c in self.centroids.items():
|
||||
dists[fam] = float(np.sqrt(np.mean((z - c) ** 2)))
|
||||
order = sorted(dists, key=lambda k: dists[k])
|
||||
if not order:
|
||||
return "", 0.0, dists
|
||||
top = order[0]
|
||||
e = {k: np.exp(-dists[k]) for k in order}
|
||||
s = sum(e.values())
|
||||
return top, float(e[top] / s if s > 0 else 0.0), dists
|
||||
|
||||
def nn(self, vec):
|
||||
"""最近邻:返回 (最像的文件名, 与它的标准化距离, 该文件所属风格)。
|
||||
|
||||
这是声线之外第二个'像不像已认可作品'的信号,且自带人耳可解释性:
|
||||
报告里能直接说出"最像 身体先听懂(urbanhiphop)"。
|
||||
"""
|
||||
z = self._z(vec)
|
||||
best = ("", 1e9, "")
|
||||
for it in self.items:
|
||||
d = float(np.sqrt(np.mean((z - np.array(it["z"], dtype=np.float64)) ** 2)))
|
||||
if d < best[1]:
|
||||
best = (it["file"], d, it.get("family", ""))
|
||||
return best
|
||||
|
||||
|
||||
def load(path):
|
||||
import json
|
||||
if not path or not os.path.isfile(path):
|
||||
return None
|
||||
with open(path, encoding="utf-8") as f:
|
||||
return StyleRef(json.load(f))
|
||||
|
||||
|
||||
def _loo_groups(items):
|
||||
"""按"歌"分组,而不是按文件。
|
||||
|
||||
坑位(踩过):最初按文件留一,得到 100% 准确率,一度以为分类器完美。
|
||||
实际上同一首歌的 _00001/_00002/_00003 是同一首曲子的不同 seed,
|
||||
音频高度同源。留一时只拿掉一个 seed,剩下 5 个"同款"还留在质心里,
|
||||
等于让样本给自己作证,准确率必然虚高。
|
||||
真正的留一必须整首拿掉——验证的是"没听过这首歌时能不能认出风格"。
|
||||
"""
|
||||
g = {}
|
||||
for i, it in enumerate(items):
|
||||
key = it.get("group") or _group_key(it.get("file", ""))
|
||||
g.setdefault(key, []).append(i)
|
||||
return g
|
||||
|
||||
|
||||
def _group_key(file):
|
||||
"""从文件名还原"这是哪一首歌"。dy10_城市晚安_swl_00001 -> dy10_城市晚安。
|
||||
|
||||
只用于分组,不是标签。
|
||||
|
||||
坑位:这里曾经把第一个下划线之前的前缀当"批次号"砍掉,只留后面的歌名。
|
||||
结果 chillhiphop_慢慢来 和 hit10_慢慢来 被并成同一组,58 个文件只出
|
||||
57 个组。留一验证时同一首歌的另一版本还留在质心里,样本等于给自己作证。
|
||||
所以只剥 seed 号和模型后缀,风格前缀必须留着——它本来就是区分同名歌的
|
||||
唯一信息。命名不规范时退回整条文件名也不会错。
|
||||
"""
|
||||
stem = file.rsplit(".flac", 1)[0]
|
||||
stem = re.sub(r"_\d{5}$", "", stem) # 去掉 seed 号
|
||||
stem = re.sub(r"_swl$", "", stem, flags=re.I) # 去掉模型后缀
|
||||
return stem
|
||||
|
||||
|
||||
def build(items, loo_min=0.65, fam_min=0.50):
|
||||
"""由带标签的语料建质心,并做按歌分组的留一验证。
|
||||
|
||||
items: [{"file","family","bpm","vec"}]
|
||||
返回可直接 json.dump 的字典。
|
||||
|
||||
关于阈值为什么这么严(实测数据,别拍脑袋改):
|
||||
语料是 58 首歌(pop 35 / hiphop 16 / punk 7)。按歌分组留一后,
|
||||
最好的组合也只到 ~65%:pop 有 4 成被判成 hiphop,hiphop 也有近
|
||||
1 成被判成 pop。逐个特征看下来没有救——tempo 单项 49%、centroid
|
||||
单项 39%,加特征、加权、改标准化都只在 60~67% 之间晃。
|
||||
根因是语料本身重叠:pop 语料里的 chill/disco/funk 变体和 hiphop
|
||||
语料在制作手法上就是连续的,不是分类器没调好。
|
||||
|
||||
所以这里刻意把门槛设高(loo_min=0.65 且每个家族都要 >=0.50),
|
||||
让 trusted 大概率为 False —— 风格判定只写进报告当参考,不当放行闸门。
|
||||
宁可报告里写"风格不可信",也不要一个 60% 准确率的分类器
|
||||
拿去做"风格不符就判 FAIL"的闸门,那是在用假精度杀人。
|
||||
风格的第一物理保证是实测 BPM 与设定 BPM 对齐(见 HARD.bpm_rel_tol)。
|
||||
"""
|
||||
import json
|
||||
X = np.array([vector(it["vec"]) for it in items], dtype=np.float64)
|
||||
med, iqr = _robust_scale(X)
|
||||
Z = (X - med) / iqr
|
||||
|
||||
cent = {}
|
||||
counts = {}
|
||||
for fam in FAMILIES:
|
||||
idx = [i for i, it in enumerate(items) if it["family"] == fam]
|
||||
if not idx:
|
||||
continue
|
||||
counts[fam] = len(idx)
|
||||
cent[fam] = np.mean(Z[idx], axis=0)
|
||||
|
||||
# ---- 按歌分组留一:整首拿掉,重算标准化和质心,再看它能不能被认对
|
||||
groups = _loo_groups(items)
|
||||
correct = 0
|
||||
total = 0
|
||||
per = {}
|
||||
confusion = {}
|
||||
for key, idxs in groups.items():
|
||||
fam = items[idxs[0]]["family"]
|
||||
keep = [i for i in range(len(items)) if i not in set(idxs)]
|
||||
Xk = np.array([vector(items[i]["vec"]) for i in keep],
|
||||
dtype=np.float64)
|
||||
mk, ik = _robust_scale(Xk)
|
||||
Zk = (Xk - mk) / ik
|
||||
ck = {}
|
||||
for f in FAMILIES:
|
||||
j = [n for n, i in enumerate(keep) if items[i]["family"] == f]
|
||||
if j:
|
||||
ck[f] = np.mean(Zk[j], axis=0)
|
||||
if len(ck) < 2:
|
||||
continue
|
||||
z = np.mean([(vector(items[i]["vec"]) - mk) / ik for i in idxs], axis=0)
|
||||
d = {f: float(np.sqrt(np.mean((z - c) ** 2))) for f, c in ck.items()}
|
||||
top = min(d, key=lambda k: d[k])
|
||||
total += 1
|
||||
good = top == fam
|
||||
correct += 1 if good else 0
|
||||
a = per.setdefault(fam, [0, 0])
|
||||
a[1] += 1
|
||||
a[0] += 1 if good else 0
|
||||
cell = confusion.setdefault(fam, {})
|
||||
cell[top] = cell.get(top, 0) + 1
|
||||
|
||||
acc = float(correct / total) if total else 0.0
|
||||
fam_acc = {k: (v[0] / float(v[1]) if v[1] else 0.0) for k, v in per.items()}
|
||||
# 总体准确率高也可能被某一个大类撑起来,所以每个家族都要单独过线。
|
||||
# 三个家族里只要有一个基本认不出来(比如 punk 全被判成 pop),
|
||||
# 这个分类器就不能拿去做风格闸门。
|
||||
fam_ok = bool(fam_acc) and min(fam_acc.values()) >= fam_min
|
||||
ref = dict(
|
||||
features=list(FEATURES),
|
||||
med=[float(x) for x in med],
|
||||
iqr=[float(x) for x in iqr],
|
||||
centroids={k: [float(x) for x in v] for k, v in cent.items()},
|
||||
counts=counts,
|
||||
items=[dict(file=it["file"], family=it["family"],
|
||||
group=it.get("group") or _group_key(it["file"]),
|
||||
bpm=it.get("bpm"), vec=it["vec"],
|
||||
z=[float(x) for x in Z[i]])
|
||||
for i, it in enumerate(items)],
|
||||
loo=dict(acc=round(acc, 4), n=total, n_songs=len(groups),
|
||||
mode="group-by-song",
|
||||
per_family={k: dict(acc=round(v[0] / max(v[1], 1), 3), n=v[1])
|
||||
for k, v in per.items()},
|
||||
confusion=confusion,
|
||||
fam_min=round(min(fam_acc.values()) if fam_acc else 0.0, 3)),
|
||||
trusted=bool(total >= 9 and acc >= loo_min and fam_ok),
|
||||
note=("由 tools/make_style_ref.py 在历史已认可歌上生成,不要手改。"
|
||||
"trusted 只有在按歌分组 LOO >= %.0f%% 且每个家族都 >= %.0f%% 时为真;"
|
||||
"实测语料 pop/hiphop 制作手法重叠,多轮尝试后总体约 60~67%%、"
|
||||
"单家族最低约 43%%,"
|
||||
"所以本文件通常 trusted=false —— 风格判定只当参考,"
|
||||
"风格的第一物理保证是实测 BPM 与设定 BPM 对齐。"
|
||||
% (loo_min * 100, fam_min * 100)),
|
||||
)
|
||||
return ref
|
||||
Reference in New Issue
Block a user