Last week, I wrote about the Articulation Index (AI) and how it helps quantify speech intelligibility. While the AI has been in use for quite some time, a more recent arrival on this scene is the Privacy Index (PI).
And here’s why I don’t like it...
First of all, the Privacy Index doesn’t add any new information. It’s calculated as 1-AI X 100 = PI (%), which is simply the opposite of the Articulation Index. So, if you have an AI of 0.30, then your Privacy Index is 70%. There isn’t any benefit to this re-representation of the Articulation Index.
The PI is also misleading. I think the problem stems from the use of the word ‘privacy’ in its name, which causes users to come to the wrong conclusion about the meaning of the rating. The fact that it’s expressed as a percentage creates even more potential for confusion. For instance, if I tell you that the Privacy Index from one place to another is 80%, you’ll likely assume you have excellent privacy – that 80% of what’s said isn’t intelligible. In reality, at 80 PI, someone can clearly understand 50% of what’s being said!
Even experienced users make this type of mistake. For example (without naming names), a major government guideline on acoustics recently quoted a sound masking vendor who said that a PI of 80 means that you understand just 20% of a conversation. The document reinforced this error by stating that a person would understand only 2 out of 10 words at a PI of 80.
So why is PI used if it’s so confusing? If I’m generous, I’d say it’s because the Articulation Index is quite abstract. Also, it was originally developed to assess how well you could understand a conversation rather than how much speech privacy you had. So it’s possible that the PI was adopted to help with these shortcomings.
However, the more cynical side of me thinks PI is used for marketing purposes. For example, a vendor may say that they’ll guarantee an 80 Privacy Index. Sounds great, but in reality it’s a low threshold. Or they may say they can improve the Privacy Index in your space by 50%, but if it’s from a low base (say 20%) then the final result is also far poorer than you’d expect.
The Articulation Index doesn’t suffer from these problems. First, most people aren’t overly familiar with the term ‘articulation’ and are, therefore, more inclined to asking about its meaning. Second, the 0 to 1 rating doesn’t strongly imply a performance level. Therefore, AI is a quantitative rating that seems more likely to spur discussion about the true level of privacy that’ll be achieved.
And what is that true level? The best method is to translate the Articulation Index into an approximation of exactly what percentage of words is understood. Interestingly, this percentage varies according to what you’re listening to. If it’s a list of individual words you already know, or a conversation you’ve heard before, then you’ll understand more than if it’s a conversation you haven’t previously heard. So, the most relevant measure would seem to be sentences upon first presentation, which best approximates a conversation you haven’t heard before. For this, there’s a well-established relationship between AI and percentage of sentences understood. While it may not be a perfect, it gives the best subjective interpretation of speech privacy levels.