Best practices
To get the most benefit from Voci's hinting feature, create hints that target frequently occurring, high-value words. Otherwise it is possible for hints to create greater confusion for the ASR engine and lower accuracy overall.
Imagine trying to solve a jigsaw puzzle. The picture of the completed puzzle on the box is a collection of hints. For example, you might see that the upper-right third of the picture is mostly blue sky, while the lower left third is dominated by a brown barn with a red door. These hints help with sub-dividing the problem into smaller problems that are easier to solve. They also serve as a cross-check on choosing the region where a particular piece belongs.
Now imagine that the picture on the box is wrong. The image is intended for a different puzzle and is therefore a misprint. The set of misleading hints in the picture would result in a significantly longer period of time required to complete the puzzle than if no hints at all were provided, due to the confusion resulting from being misdirected.
The impact on Voci transcription is similar, but rather than taking longer, the misdirection would result in less accurate transcripts.
Consider the following best practices to be guidelines rather than absolute rules. Hints that go against these guidelines may still be useful. For example, the hint "Frigidaire refrigerator" may appear to be over-constrained, but it would be a useful hint if this word pair is known to occur with high frequency in the call population. It is acceptable for a hint to be a subset of another hint.
Categorize hints correctly
In the hardware store example, consider that how people speak about company_name words is significantly different from how they speak about brand_name words. Misclassifying a term can result in the ASR engine looking for phrases such as "Thank you for calling Tilex" or "I bought a Joe's Hardware lawnmower". These phrases are unlikely to be spoken and are therefore misleading.
Properly constrain hints
A good hinting phrase contains only the words necessary to capture the essence of the hint. For example, if the company name is "Joe's Hardware" and these two words are used together frequently, then the proper hinting phrase is "Joe's Hardware". An under-constrained set of hints would be "Joe's" and "Hardware" as separate hint phrases. An example of an over-constrained hint is, "Joe's Hardware how may I help you".
In the under-constrained case, recognition of "Joe's" and "Hardware" will be increased independently, but recognition of the phrase "Joe's Hardware" will not be improved significantly. Also, because "Joe's" and "Hardware" are short words, the number of false positive occurrences of these words could increase. Shorter hint phrases are more easily confused with speech that includes sounds similar to those in the hints, resulting in false positives.
In the over-constrained case, it would be better to remove "how may I help you" from the hint, because this variant is only one example of a phrase that frequently follows "Joe's Hardware". Including the extra context unnecessarily constrains the expected usage of "Joe's Hardware", which will reduce the effectiveness of the hint when the context is different.
Keep number of hints to a minimum
Resist the temptation to create large numbers of hints, especially within a single category (e.g. brand_name).
Keep hints focused on high-value phrases that are known or expected to be transcribed with degraded accuracy in the absence of hints.
Where possible, use available metadata to further narrow down the set of hints likely to be applicable to any given call. For example, in a voicemail use case the caller's phone number could be matched against a name in the receiver's contact list.
In dynamic hinting, an advanced use of hinting, a script is used to generate a set of hints which include caller name and phone number. The dynamic hint is used while the voicemail is transcribed. This example relies upon a more advanced method of deploying hints, but dynamic hinting is worth mentioning here because this approach illustrates how the number of hints applied to any given transcription session can be decreased to only those likely to provide value. Advanced uses of hinting should be discussed with Voci support.
