Improved Tonal Language Speech Recognition by Integrating Spectro-Temporal Evidence and Pitch Information with Properly Chosen Tonal Acoustic Units.