摘要
在音频分类中，参数较少的可微听觉滤波器组覆盖了硬编码频谱图和原始音频之间的中间地带。LEAF（arXiv:2101.08596）是一种基于Gabor的滤波器组，结合了每通道能量归一化（PCEN），已经显示出有希望的结果，但计算成本较高。通过使用非均匀卷积核大小和步幅，并用更好的并行化操作替代PCEN，我们可以更高效地达到类似的结果。在六项音频分类任务的实验中，我们的前端以LEAF 3%的成本达到了相同的准确率，但两者都无法始终超越固定的梅尔滤波器组。可学习音频前端的研究尚未得到解决。
代码仓库
cpjku/efficientleaf
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| audio-classification-on-birdclef-2021 | melspect | Accuracy: 39.9 |
| audio-classification-on-birdclef-2021 | EfficientLEAF | Accuracy: 42.9 |
| audio-classification-on-birdclef-2021 | LEAF | Accuracy: 42.3 |
| audio-classification-on-birdclef-2021 | EfficientLEAF (8s) | Accuracy: 72.2 |
| audio-classification-on-crema-d | LEAF | Accuracy: 50.2 |
| audio-classification-on-crema-d | melspect | Accuracy: 58.8 |
| audio-classification-on-crema-d | EfficientLEAF | Accuracy: 60.2 |
| audio-classification-on-speech-commands-1 | melspect | Accuracy: 95.1 |
| audio-classification-on-speech-commands-1 | EfficientLEAF | Accuracy: 95.2 |
| audio-classification-on-speech-commands-1 | LEAF | Accuracy: 95.1 |
| instrument-recognition-on-nsynth | EfficientLEAF | Accuracy: 71.7 |
| instrument-recognition-on-nsynth | LEAF | Accuracy: 69.2 |
| instrument-recognition-on-nsynth | melspect | Accuracy: 72.1 |
| spoken-language-identification-on-voxforge-2 | LEAF | Accuracy: 91.5 |
| spoken-language-identification-on-voxforge-2 | EfficientLEAF | Accuracy: 86.6 |
| spoken-language-identification-on-voxforge-2 | melspect | Accuracy: 85.6 |