Speech Enhancement Auto-Encoder

Network Architecture:

The auto-encoder part is borrowed from SEGAN's generator.

Results:

	PESQ	SSNR
Noisy	1.97	1.68
SEAE	2.2	7.6

Measurement implementation is included in Speech Enhancement: Theory and Practice. Publisher website.

Dependencies:

• Tensorflow 1.14

Data:

• Please download the Dataset (Valentini et al.2016).

• Unzip and move the directories into data.

• Run python tfrecords.py

Format : 16kHz sample rate, 16bit signed integer

Usage:

Train

./train.sh&

Training results (trained model, log) are saved in results/.

Inference

./inference.sh

You can find the enhanced data in inferences.