nlp-deeplearning-club / kaggle-project Goto Github PK

参考:

• https://www.kaggle.com/alexozerin/end-to-end-baseline-tf-estimator-lb-0-72

https://github.com/buriburisuri/speech-to-text-wavenet

先这么多

主要是为了满足树莓派那边的需要

搞明白咋回事,关键是用了试试.

https://github.com/rizar/attention-lvcsr

https://github.com/castorini/honk

简单，迅速浏览了一下现有的 Kernels，分类不一定准确，可以之后矫正。

1 基于波频特征提取的：
1a 大概提取了波频最大最小值等等特征
https://www.kaggle.com/the1owl/a-tf-will-emerge/code
2b 随机森林
https://www.kaggle.com/lucifer19/speech-recognition-challenge-3d-waves-sg

2 基于将波形转换到频率域的，类似变成图像，然后CNN：
https://www.kaggle.com/timolee/audio-data-conversion-to-images-eda
https://www.kaggle.com/voglinio/keras-directory-iterator-wav-to-spectogram
https://www.kaggle.com/solomonk/pytorch-speech-recognition-challenge-wip
https://www.kaggle.com/sainathadapa/keras-starter-code

3 基于傅里叶变换和信息熵
https://www.kaggle.com/solomonk/single-wav-fft-entropy

https://github.com/zzw922cn/Automatic_Speech_Recognition

在数据预处理之前做一步数据增强,主要是

1. 确保提取主干没问题
2. 加噪音
3. 声波能量强度整体上的拉伸压缩
4. 频移
5. 时域上的拉伸压缩

rt

搞明白咋回事,关键是用上,主要要看的就是attention-is-all-you-need这篇和它的实现

参考

• https://www.kaggle.com/davids1992/speech-visualization-and-exploration
• https://www.kaggle.com/timolee/audio-data-conversion-to-images-eda
  +https://www.kaggle.com/solomonk/single-wav-fft-entropy

https://github.com/mozilla/DeepSpeech