ShiyiLIN0922/WildEar-HZ-strict-005

WildEar-HZ strict-005

中文在前，English follows.

中文

模型简介

WildEar-HZ strict-005 是一个轻量生物声学后训练包。它不包含 BioLingual 基础模型权重，而是包含：

• 21 类目标物种原型；
• 21 类文本原型；
• 8 类非目标动物原型；
• Unknown 双阈值；
• 固定版本的分类、交叉验证和独立 open-set 评估；
• 可重放的训练与评估脚本。

运行时使用 davidrrobinson/BioLingual 冻结音频编码器生成 512 维单位嵌入。

输入与输出

• 输入：48 kHz 单声道音频；
• 分窗：5 秒窗口、2.5 秒步长；
• 最大输入：60 秒；
• 输出：21 个目标类别中的三个排序候选、模型参考时间片段和 Unknown 风险提示。

候选内显示分数是温度归一化的相对分数，不是校准后的真实后验概率。

训练方法

许可弱标签音频按上传者或来源组全局隔离。高文本 margin 窗口用于构建类别原型：

p_c = normalize(mean(z_i))

题目方明确标签声音构建参考原型：

p'_c = normalize((1 - alpha) p_c + alpha r_c)

alpha=0.15 仅由 validation 选择。21 类中有 13 类使用训练音频原型，8 类使用文本回退。240 条野外候选中的 16 条 AI 伪标签未通过固定 validation 门禁，生产采用数为 0。

最终来源隔离 test

• 305 train / 111 validation / 88 test；
• 三组上传者交集为 0；
• test：88 条录音、43 位独立上传者、17/21 类；
• Accuracy：90.91%；
• Macro-F1：81.43%；
• 前三候选命中率：95.45%。

嵌套 5 折上传者分组交叉验证

范围：454 条录音、13 类、每类至少 5 个独立来源组。

方法	Accuracy	Macro-F1	前三候选
BioLingual 零样本	88.11%	83.94%	96.26%
类别原型后训练	94.05%	91.95%	99.12%
Class-balanced 线性头	94.71%	93.09%	99.34%

原型后训练相对零样本的上传者组 Bootstrap 提升：

• Accuracy：中位数 +5.78 个百分点，95% 区间 +0.93 至 +11.67；
• Macro-F1：中位数 +8.11 个百分点，95% 区间 +1.40 至 +15.36。

线性头相对原型法的区间跨 0，因此当前证据不支持显著优势。生产版本保留类别原型。

独立 open-set test

非目标动物按上传者全局拆分；环境声按 ESC 原始 fold 拆分。非目标原型训练、阈值选择和 test 交集为 0。

分组	AUROC	线上 Known recall	线上 FAR
总体	0.9696	72.73%	6.45%
非目标动物	0.9500	72.73%	20.00%
环境声音	0.9789	72.73%	0.00%

Unknown 是保守风险提示，不替代专家物种鉴定。

文件

• wildear_biolingual_prototypes.npz：目标与文本原型；
• wildear_unknown_animal_prototypes.npz：非目标原型与线上阈值；
• wild-ear-hz-strict-005.json：release manifest 与 SHA-256；
• reports/：最终 test、嵌套分组 CV、独立 open-set test；
• TECHNICAL_REPORT_ZH_EN.md：中英双语技术报告；
• TECHNICAL_REPORT_ZH_EN.pdf：中英双语详细 PDF。

使用限制

1. 最终标签为平台录音级弱标签，不是专家事件级金标准。
2. 4 类未进入最终 test。
3. 8 类仍使用文本回退。
4. 当前主链不是正式多标签事件检测。
5. 非目标动物 FAR 高于环境声。
6. 不应用于受保护物种执法、种群数量估计或生态影响评价。

---

English

Overview

WildEar-HZ strict-005 is a lightweight bioacoustic post-training package. It does not redistribute BioLingual base weights. It contains:

• 21 target-species prototypes;
• 21 text prototypes;
• eight non-target-animal prototypes;
• dual Unknown thresholds;
• fixed classification, cross-validation, and independent open-set reports;
• replayable training and evaluation scripts.

At runtime, the frozen davidrrobinson/BioLingual encoder produces normalized 512-dimensional audio embeddings.

Input and Output

• Input: 48 kHz mono audio;
• windows: 5 seconds with a 2.5-second hop;
• maximum duration: 60 seconds;
• output: three ranked candidates from 21 target classes, model-reference time segments, and an Unknown risk signal.

Displayed candidate scores are temperature-normalized relative scores, not calibrated posterior probabilities.

Training

Licensed weak-label audio is globally separated by uploader/source group. High text-margin windows form class prototypes:

p_c = normalize(mean(z_i))

Explicit challenge references form adaptation prototypes:

p'_c = normalize((1 - alpha) p_c + alpha r_c)

alpha=0.15 is selected only on validation. Thirteen of 21 classes use audio-trained prototypes; eight use text fallback. Sixteen AI pseudo-label candidates fail the fixed-validation gate, so production uses none.

Final Source-Grouped Test

• 305 train / 111 validation / 88 test;
• zero uploader-group overlap;
• 88 test recordings from 43 uploaders, covering 17 of 21 classes;
• 90.91% Accuracy;
• 81.43% Macro-F1;
• 95.45% top-three hit rate.

Nested Five-Fold Grouped Cross-Validation

Scope: 454 recordings from 13 classes with at least five source groups per class.

Method	Accuracy	Macro-F1	Top-three
BioLingual zero-shot	88.11%	83.94%	96.26%
Prototype post-training	94.05%	91.95%	99.12%
Class-balanced linear head	94.71%	93.09%	99.34%

Uploader-group bootstrap prototype-over-zero-shot gains:

• Accuracy: median +5.78 points, 95% interval +0.93 to +11.67;
• Macro-F1: median +8.11 points, 95% interval +1.40 to +15.36.

Linear-head-versus-prototype intervals cross zero, so production retains prototypes.

Independent Open-Set Test

Non-target animals are globally separated by uploader; environmental audio is separated by original ESC fold. Prototype-train, calibration, and test overlaps are zero.

Group	AUROC	Online Known recall	Online FAR
Overall	0.9696	72.73%	6.45%
Non-target animal	0.9500	72.73%	20.00%
Environmental sound	0.9789	72.73%	0.00%

Unknown is a conservative risk signal, not an expert species identification.

Limitations

1. Final labels are platform recording-level weak labels, not expert event annotations.
2. Four classes are absent from final test.
3. Eight classes still use text fallback.
4. The main task is not formal multi-label event detection.
5. FAR is higher for non-target animals.
6. Do not use this model for enforcement, abundance estimates, or ecological-impact assessment.