When reading out the same order number, different people say different things.
For example, #1002-7899
might be read as
… or a vast number of other ways.
This poses a challenge in automated conversations in contact centres. We came up against it in a recent project with a retail client.
This is how we solved it:
In the Dialogflow Intent, we used the system entity @sys.number-sequence. This captures the ‘dash’.
Then we implemented a code logic in Dialogflow Fulfillment to remove the ‘dash’ and only process the numbers.
Since we were using Google’s STT (speech-to-text) service alongside Twilio, we were able to improve the accuracy with Twilio’s speech hints. This allowed us to bias STT towards certain phrases depending on context.
These 'order number intents' could also be configured in Twilio with hints like 0
1
2
3
4
5
6
7
8
9
double
triple
to get better STT results.
Adding Twilio’s hints increased our accuracy from 90% to 95%.
Let us know if you’ve come up against this challenge before.