Now it works in my ESL app though I am just able to do one dialogue ( I need to add the event catching for furthur dialgoues).

I have a couple of questions here:

1. In the first try, my Nuance server was able to be accessed somehow (FS says the MRCP is not responding in 5000ms,
    something like that), then FS says: [WARNING] rtsp_client.c:386 () Failed to Connect to RTSP Server MailScanner warning: numerical links are often malicious: 99.185.85.31:554,
    later FS says:
     [ERR] mod_unimrcp.c:1860 (TTS-6) SYNTHESIZER channel error!
     [ERR] switch_ivr_play_say.c:2439 Invalid TTS module!

   The SYNTHESIZER channel error and Invalid TTS module error are obvious.

    What I don't understand is why it went to this stange address: MailScanner warning: numerical links are often malicious: 99.185.85.31:554?

check your unimrcp configuration. Make sure the default TTS and ASR profiles are set to actual servers.

2. I specified TTS engine in play_and_detect_speech as
         "say:unimrcp:nuance5-mrcp1-1: the text to speak"
     It works though I didn't specify the TTS voice.

     How do I specify the TTS voice? In the mrcp profile (how?)? or something like:
         "say:unimrcp:nuance5-mrcp1-1:Serena: the text to speak" (this seems not right.)

That won't work. Set the tts_engine variable as I explained previously, or use say:unimrcp:voice:text to speak with the desired voice and the correct default TTS profile defined in unimrcp.conf.xml. This is a limitation of the say: notation. Alternatively, the voice can be defined with the tts_voice channel variable.

3. The barge-in works well, thanks!. Is the barge-in configurable? In some scenarios, we might not allow barge-in.

If you don't want to barge in, just do "playback (or speak)" first, then "play_and_detect_speech" with a silence prompt.

4. How could I get the text which has spoken to the user when barge-in occurs?
Or Could I get the time when barge-in occurs? If I know the barge-in time and rough totale time for the whole text
to be spoken I can figure out the spoken text by manually checking the recorded audio file later, which would be painful.

If this is necessary, you might want to use the lower-level functions instead to watch for the begin-speaking event.

5. when I use "speak" and "detect_speech" apps in ESL, I can catch event: DETECTED_SPEECH and speech-type: begin-speaking
     and "detected-speech", then I do the recognition results processing.

    The new app play_and_detect_speech seems not generate these events any more. The way that I can think of to get the results
    is to catch event:CHANNEL_EXECUTE_COMPLETE then check if variable_current_application=play_and_detect_speech, then get
    the results from variable_detect_speech_result.

    Is this the proper way to get the results in ESL app? Or will play_and_detect_speech later on be consistent with detect_speech
    in term of ASR events?

play_and_detect_speech is a higher level abstraction to simplify things. If you want to have more control, go back to using the ESL events. Reading the code in mod_dptools and switch_ivr_async will give you hints about how to do it correctly.

6. I'd like to set start-input-timers=false in the initial request then start the recognition timers (start-input-timers=true)
after the TTS finishes.
How possibly could I do this?

This is automatically done in the switch_ivr_play_and_detect_speech() function. You just need to specify start-input-timers=false in the beginning.