We are making a Lex bot V1 integrated with Amazon connect for the purpose of booking a cab. One of the requirements is to get an address from the user using a single prompt. Lex provides a built- in slot called Amazon.StreetName for getting street addresses from users. We have tried using it for our bot, and it seems to work fine with user address utterances like "eight chapel hill" or "fifteen cross lane" but only in the lex testbot console.
When talking to bot over call, using the number claimed in connect, the bot's transcribe / Speech to text performance gets really bad and it is not able to transcribe anything user utters for addresses, except for yes/no confirmations.For complex addresses like user said "tesco kirkdale" bot takes as "has girl deal"(sample attached - 6). Although, slow and clear user utterances improves transcription quality but bot still captures some noise in input, and also we cannot expect our users to be articulating words to the bot all the time. Performance of the bot is similar for a custom slot type "Address", we have defined for taking pickup or drop off addresses. Do we need to make any custom vocabulary with corresponding pronunciation in S3, for fine tuning the AWS lex transcribe, if yes, how can we do that?
Bot is also not able to understand AMAZON.number slot utterance for "one” which is transcribed as "juan"/"jone".
We also have problems with some non english names like "Ifti", bot takes it as "b50" or "if the" or "Momal" which it takes as "Nom"etc.
Using custom slot for taking addresses or AMAZON.StreetName, when user utters an address in format of "street number street name", the number in the beginning of the utterance also triggers the wrong slot i.e. AMAZON.Number(slot name - “Passengers”, slot type- :AMAZON.Number) and takes number for passenger slot and street name for address slot(Custom slot type -Address, slot name:“PickUpLocation”). Although the slot priority is defined. For instance, if the user utters "nine chapel hill/forty three bradford street" , bot transcribes as "yes oh nine o res" /"forty three bradford street" (sample attached - 4), considering it as AMAZON.Number slot value.
how can we Improve amazon lex bot speech to text for addresses?
