##### Manually Identifying Errors

The following sections focus on indirect search tactics that have proven useful for locating good substitution candidates. Perform these manual searches to locate substitution candidates. Then, validate errors by listening to the associated portion of audio. Once the errors have been validated, you can develop a list of transcription errors and their corrections.

###### Positional Search

Some conversation types, such as customer service calls, have opening and closing statements that appear in the majority of conversations. These recurring statements with their common positions are good places to start looking for substitution candidates. Search near the beginning and end of transcripts for familiar words and phrases. For example, most call agents will greet callers saying "Thank you for calling <company name>, how may I help you?" In this instance, examine the agent's greeting in a few transcripts to see if the company's name was transcribed correctly.

For example, agents in the call center for "Wells Fargo Foreign Exchange" frequently answer calls using the company's name. If you review the beginning of several transcripts of these calls, you may discover that the company name is being transcribed as "Wells Fargo for a change". This mistranscription is a good candidate for substitution. The following substitution rule would correct the transcription error.

wells fargo for a change : /Wells/ /Fargo/ /Foreign/ /Exchange/

Substitution rule syntax is explained in detail in Substitution Syntax.

Positional searching can also be used to find substitution candidates in scripted speech. Call agents will often use scripted speech such as disclaimers, company descriptions, product descriptions, transaction confirmations, and other constructs. These distinct statements will stand out in your transcripts and are likely to appear in the same area of each transcript. Review the scripted speech in some transcripts to verify that the statements are transcribed correctly.

Use positional searching on areas where you know what might be said, such as the beginning and end of transcripts. Substitution candidates will be easiest to identify in these positions because you will be most familiar with the transcription results.

###### Contextual Search

You may suspect certain words or phrases of being mistranscribed. Performing a contextual search is a good approach in this case. Search for the context surrounding the word or phrase you are trying to verify. Some guesswork is involved in contextual searching. Therefore, an understanding of how the target word or phrase typically gets used is a prerequisite to performing a contextual search.

The words or phrases suspected of being mistranscribed will often be preceded or followed by predictable phrases. For example, the phrase “credit card” or "PayPal" can usually be found near “payment”, “make a payment”, or “how will you be paying”. Search for words or phrases that you expect to be in close proximity to the word or phrase you are looking for. Then, listen to the associated portion of audio to verify the suspected words or phrases. The transcription error that is corrected in the following substitution rule was found by searching for "Nissan Motor" instead of the entire phrase.

Nissan Motor except this Corporation : /Nissan/ /Motor/ /Acceptance/ /Corporation/

### Tip

If you suspect a known phrase of being mistranscribed, try searching for a portion of that phrase.

###### Phonetically Similar Words and Phrases

Look for words that sound similar to words in your target phrase. For example, "date of birth" could be mistranscribed as "data earth." The following example contains a list of substitution rules. Each of these rules have a similarly sounding "before" and "after" pair.

current see : currency
been fishery : beneficiary
data earth : date of birth

Read through your transcripts carefully to find instances of words or phrases that are frequently mistranscribed as phonetically similar words or phrases. Similar sounding words are good substitution candidates and become easier to predict with practice. When searching for phonetically similar words and phrases, try to think of unusual words (or parts of words) that sound like the target word or phrase.

###### Unexpected Words and Phrases

Look for words or word patterns that are unlikely to be spoken, either in general or in cases specific to your business domain. These variances are often associated with errors.

For instance, three-letter words that end in "re" are often associated with errors because there aren't many three-letter words that end in "re" (with "are" being a notable exception). Using the V‑Spark search tool, specify the regular expression “[^’a]re” to search for such words. V‑Spark will return all three-letter words that end in "re" where the first character is not an "a" or an apostrophe. This constraint will prevent words like "we're" from appearing in search results. The apostrophe acts as a word boundary making the search engine think "re" is a whole word rather than the end of a longer word.

To find more specific substitution candidates, look for words that appear out of context or that are irrelevant to your industry. For example, the word "whale" isn't likely to be spoken in calls specific to auto insurance companies. If "whale" appears in transcripts related to auto insurance, this word is probably a candidate for substitution. The associated audio portion would then be reviewed to reveal a customer filing an insurance claim for damage done by "hail" and not "whale".

The following example illustrates substitution rules that correct unexpected words and phrases.

whale damage: hail damage
hit a beer : hit a deer
pay leo : paleo
###### Manual Scanning

Manual scanning requires you to listen to calls and read their associated transcripts. When you find substitution candidates, use V‑Spark's search capabilities to find other instances of the same candidate, then listen to the associated portions of audio to verify suspected phrases. Manual scanning is time consuming but this process will enable you to locate errors not found by other methods.