Substitution

Substitution works after speech-to-text transcription has completed, correcting errors in transcripts automatically using rules that find and replace errors with corrected text.

Transcription errors occur for a variety of reasons. Typical reasons for transcription errors include:

• Words that are not included in the language model dictionary

• Fast or poorly enunciated speech

• Unlikely word associations or industry-specific terms

• Poor audio quality

The goal of transcript analysis is to extract actionable insights to make more informed decisions. Transcription errors hinder that analysis because poorly transcribed words and phrases can limit or skew the results. Substitution is an effective transcription-tuning approach when used to target the words and phrases that directly affect transcript analytics.

Substitutions operate best on errors where specific words or phrases are frequently mis-transcribed in the same or similar ways. A before : after pair defines the error and the correct replacement. For example, if the spoken phrase "date of birth" is transcribed frequently as "data birth", this can be corrected with a substitution rule of data birth : date of birth.

Substitution rules

Analyze the transcripts to locate errors. Listen to call audio to confirm that the same word or phrase is being spoken for any suspected transcription error. After you have identified errors, you can develop substitution rules to improve transcription accuracy.

Incorporate substitution rules into the query sent to the bulk ingestion API to trigger audio file processing. See the example below for proper formatting. By default, rules are not case sensitive, so capitalization of a replacement phrase will match the capitalization of the original phrase. You can, however, control letter case by enclosing a word between forward slashes. Each word in the replacement text must be delimited individually, as shown in the last rule of the following example:

{"n d a" : "/NDA/", "pc and number" : "/PCN/ Number", "it's vance physical therapy" : "/Advanced/ /Physical/ /Therapy/"} 

Custom vocabulary

Custom vocabulary (also known as out-of-vocabulary or OOV) is an ASR tuning feature for the Voci engine (Engine 1) that improves transcription accuracy for audio that contains brand- and industry-specific terms. Custom vocabulary supports all Voci language models and enhances them with preferential treatment for new words.

For more information, see Custom vocabulary.

Redaction

For security, Speech automatically redacts credit card numbers, social security numbers, and street numbers from transcripts and audio. You can choose to disable redaction if you wish to keep that information visible in Experience Cloud. To do so, set apply_redaction to "No" in the transcription API request.

Speech's redaction algorithm uses simple regular expression patterns to redact numbers. By default, all numerals are redacted, with the exception of the following patterns:

• Ordinal numbers
• Percentages
• Times
• Prices
• Short decimal numbers

Single-digit numbers

Speech tends to transcribe single-digit numbers into a word and not a numeral. For example, the utterance "eight" is transcribed as eight instead of 8. If a spoken number is transcribed as a word, that word is not redacted. As a result, single-digit numbers are typically not redacted. If a spoken number is transcribed as a numeral, it is redacted by default.

Credit card numbers

Speech uses the Luhn algorithm to identify credit card numbers in transcripts. When redaction in on, the Luhn presence field (a_media_luhn) identifies whether a redacted credit card number is present in a transcript. For example, some companies require that call center agents turn off voice recordings when customers share their credit-card numbers. When the Luhn presence field has a value of "Yes", the agent for that call did not turn off voice recording.

To learn more about how redaction works, see Redaction syntax and default rules.