Created by: xinghai-sun

• Prepare one or more internal Mandarin speech recognition datasets for our internal benchmark.
• If there is any particular data preprocessing for Mandarin, add it to the audio data provider.
• Prepare a reliable baseline & evaluation details. It would be better if we could try the baseline training environment ourselves.
• Need cooperating with the Department of Speech in Baidu.
• Refer to the DS2 design doc and update it when necessary.