Transcription & Phrase Spotting Improvements in Invoca

[Jacaselon]

Contents of this Article

• Transcription
• Transcription Details
• Phrase Spotting
• Phrase Spotting Details

Transcription

This document provides an overview of how our transcription system works. It highlights the differences in how spoken phrases are interpreted, focusing on elements such as time formatting, percentages, URLs, and hyphenated words from real-life conversations. Generally, transcripts format the translated text as the words are spoken.

We have made a few updates to improve the overall experience with Transcription in Invoca to ensure the calls are being transcribed as efficiently and effectively as possible. These nuances and improvements are displayed below.

Category	General Transcription	Transcription in Invoca
Numbers	"One, two, three"	"123"
Currency ("bucks")	"One Buck, 34 Bucks"	"$1, $34"
Currency ("dollars")	"20 dollars"	"$20"
Time	"630PM"	"6:30pm"
Percent	"fifty percent"	"50%"
URL	"liberty mutual dot com"	"libertymutual.com"
Hyphenated words	"autopay"	"auto-pay"
Brands	“United healthcare”	“UnitedHealthCare”
Slash (“/”)	“Home depot slash landing”	“Homedepot/landing”

Transcription Details

Pauses in Read-Outs

In cases where customers are reading out an email or a number, such as a phone number or credit card number, a slight delay between readouts can cause the number to be divided within the transcript.

Number Pronunciation

When customers read out lengthy numbers such as credit card, account numbers, or phone numbers, variations in pronunciation can lead to inaccuracies in transcription. For example, "4050" might be read as "Forty Fifty" or "Four Oh Five Oh."

Phrase Spotting

Our improved transcription capabilities also include enhanced phrase spotting. The system formats text as it would be displayed in print. For example, it might produce "$100" instead of "one hundred dollars" or "invoca.com/landing" instead of "invoca dot com slash landing."

To ensure accuracy, we first attempt to match the words as they appear in the transcript. If that doesn't work, we reverse the formatting and try matching again. This dual approach helps maintain precision in transcription matching.

The phrase spotting UI now accommodates special characters to enhance various use cases.

Category	General Transcription	Transcription in Invoca
Hyphens	“Checkin”	“Check-in”
Percentage symbol	“Sixty percent”	“60%”
Currency symbols	“Sixty dollars”	“$60”
Commas and periods	“Red white and Blue”	“red, white, and blue” or “red. white. And Blue.”
@ Symbol	“Janedoe at gmail dot com”	“Janedoe at gmail.com”
Ampersand	“AT and T”	“AT&T”
Periods within word phrase	"Forty dollars and Fifty cents" or “Aol dot com”	“$40.50 and/or AOL.com”
Numerization	“Five” or "Fifty"	“5 or 50”
Numerization with commas	"Twelve Hundred" "One Thousand Two Hundred" "One Million Five Hundred Thousand"	“1,200” or “1,500,000”
Numerization with hyphens	“One Thousand One Hundred”	“1,100”
Years	“Two thousand twenty” or “nineteen ninety five”	“2020” or “1995”
Slash in words	“proposal estimate”	“proposal/estimate”
Slash and hyphen in words	“Next night”	“next-night”

 

Phrase Spotting Details

Flexible Matching

Phrase lists can be configured to match phrases as seen when viewing a transcript. Generally, a phrase matches a transcript exactly, but there are cases where we allow the match to be less strict:

• Word Stemming: When comparing words, we use their root forms if available. For instance, the word "walk" would match with "walked" and "walking" in the transcript, and vice versa.
• Number Partial Matches: When matching numbers, if the phrase consists solely of a number, it can match numbers in the transcript that include symbols. For example, the phrase "100" would match transcripts containing "$100" or "100%." However, the reverse is not true. A phrase with a number and a symbol will only match transcripts with the exact same number and symbol.
• Slashes and Dashes: Slashes and dashes are treated as spaces when comparing phrase lists and transcripts. Therefore, the phrase "this-that" would match transcripts containing "this-that," "this that," or "this/that."