import numpy as np
from soundscape.features import FeatureConfig, logmel_spectrogram
cfg = FeatureConfig(sample_rate=32000, n_fft=1024, hop_length=320, n_mels=64)
signal = np.random.default_rng(0).standard_normal(cfg.sample_rate)
feats = logmel_spectrogram(signal, cfg) # shape: (帧数, 64)FeatureConfig 是不可变的 dataclass，frame_rate（帧率）与 effective_fmax
（有效上限频率）会自动推导。
分类器工作在 clip 级嵌入上，需要先把帧级特征池化成一个向量：
from soundscape.asc import PrototypeSceneClassifier, mean_std_pool
X = np.vstack([mean_std_pool(logmel_spectrogram(s, cfg)) for s in train_signals])
clf = PrototypeSceneClassifier().fit(X, train_labels)
pred = clf.predict(mean_std_pool(feats)[None, :])
proba = clf.predict_proba(mean_std_pool(feats)[None, :])检测器需要一个满足 FrameModel 协议的帧级模型（features → 每帧每类概率）。
仓库自带一个线性基线，方便先把链路跑通：
from soundscape.sed import LinearFrameModel, SoundEventDetector
from soundscape.sed.labels import DESED_EVENTS
model = LinearFrameModel.random(cfg.n_mels, len(DESED_EVENTS))
detector = SoundEventDetector(model, DESED_EVENTS, cfg, threshold=0.5, median_kernel=5)
events = detector.detect(signal) # list[Event]替换成自己的模型时，只要它接受 (帧数, 频带) 并返回 (帧数, 类别数) 概率即可。
from soundscape.metrics import SegmentBasedMetrics, EventBasedMetrics
seg = SegmentBasedMetrics(DESED_EVENTS, time_resolution=1.0)
seg.evaluate(reference_events, estimated_events)
print(seg.results()["overall"]) # precision / recall / f_measure / error_rate
evt = EventBasedMetrics(DESED_EVENTS, t_collar=0.2)
evt.evaluate(reference_events, estimated_events)
print(evt.results())from soundscape.llm import SceneNarrator
print(SceneNarrator(language="zh").describe("park", events))
from soundscape.retrieval import EmbeddingIndex, search
index = EmbeddingIndex(audio_embeddings, ids=clip_ids)
similarity = index.similarity(text_embeddings)soundscape scenes
soundscape events
soundscape describe --scene park --events est.txt --lang zh
soundscape evaluate --reference ref.txt --estimated est.txt --mode segment
soundscape demo --seconds 4 --seed 0