Virtual Agent and TTS - Best Practices

Table of Content:

IVR Script:

1. Know your voices:

1. Give the customer a sample start script with all voices possible, separating Wavenet and Standard voices, with commercial impact added.

2. Here’s how to download voice prompt: https://drive.google.com/file/d/15DjeXjODe9MOF-JXV-E2I_2Efwk_s2Ar/view?usp=sharing

3. Remember, static prompts (prompts which do not have any dynamic text in it) is only generated once and re-used without incurring new charges.

2. Start with:

1. Add a connection delay. We recommend adding 500 ms to 1 sec delay. This delay will change based on Inbound or Outbound modes.

2. Start with Greetings, such as “Good Morning”

3. Add the customer name whenever possible, eg: “Good Morning, Abhishek”

3. Script:

1. Use SSML Tags

2. Use Grammarly or any other service to check the grammars first.

3. Consider having a single DTMF number to go back one level (“Press 8 to go back”) and one for going back to Main Menu (“Press 9 to go to the Main Menu”)

4. Use SSML Tags such as

• Go through all SSML Tags from here: https://cloud.google.com/text-to-speech/docs/ssml#

• <emphasis> to emphasise important bits in the script. Keep the emphasis on one word only.

• Use <break time='500ms'> to add break after every sentence or before important informations.

  • 1500ms between paragraphs

  • 500ms-800ms between sentences.

  • 400ms+ for any other place.

• Use <s>...</s> tags to wrap full sentences, especially if they contain SSML elements that change prosody (that is, <audio>, <break>, <emphasis>, <par>, <prosody>, <say-as>, <seq>, and <sub>).

• If a break in speech is intended to be long enough that you can hear it, use <s>...</s> tags and put that break between sentences.

5. How to ask questions:

• Questions should be short and sweet

• Ask to the point and close ended questions

6. DTMF options:

• Always configure the “Length” of DTMF input to ensure that after correct. Most of the time you’ll be taking a 1 digit input (i.e. 1 to 9), so please ensure Length is set to 1.

• Add pause after important bits, example: Press 1, for account information to add a delay after “Press 1”

  • Add break delay after each option for enough time for customer to process the information. We recommend <break time=”600ms”>

• Always add an option to repeat the information.

Virtual Agent Training

DialogFlow Training

• As important as the flow is the Intent Training

• Create the required intents in Google DialogFlow or IBM Watson

• Create lots of Entities!!

• Reference:

  • https://medium.com/janis/9-ways-to-make-your-dialogflow-agents-more-reliable-2b455f1a5b8b

  • https://cloud.google.com/dialogflow/docs/training

• Handle All Branches

• Make sure one covers all the scenarios and branches of the widgets

Single Utterances

• Single utterance is the function of Google which helps to identify single word inputs in a better way

• Enabling single utterances is not a sure shot solution in case of languages other than Eng(US) as none of them have the phone call model by Google enabled for STT

• One can enable single utterance and choose not to add any “Expected Phrases” if not needed

• To know more check out the user manual /wiki/spaces/PD/pages/17137699

• Expected Phrases

  • Class tokens as well as single utterance words, phrases can be added under expected phrases to let the system understand and better detect the speech

  • Make sure no “Space” is added between two expected phrases i.e. Correct → Yes,Yes this is correct and Wrong → Yes, Yes this is correct

  • To know more check out the user manual /wiki/spaces/PD/pages/17137699