Whether your goal is to create a high degree of comprehension or a high degree of privacy, it’s helpful to have a standard way of measuring the level of speech intelligibility provided in a space.
This week, I’d like to address one of these – the Articulation Index (AI) – which was developed at Bell Labs in 1921 by one Harvey Fletcher during his efforts to quantify speech comprehension over telephone lines. During the 1950s, those involved in the speech privacy sciences adopted Harvey’s invention as a measure of exactly the opposite: how much one couldn't understand.
The Articulation Index is calculated based on measurements of how loud a conversation is relative to the background sound levels in a space. As I mentioned in an earlier blog, this is known as the signal-to-noise ratio. For the AI, the signal-to-noise is measured in each of twenty narrow frequency ranges (from 200 to 5,000 Hz). Each of these is then weighted according to the degree to which each frequency contributes to comprehension of speech. The total is added to arrive at the Articulation Index, which is expressed as a numeric rating ranging from 0 (no comprehension) to 1 (perfect comprehension).
Curiously, while the Articulation Index was developed to assess speech intelligibility, it doesn’t directly tell you how much a person will understand. Further experiments had to be conducted in order to connect actual levels of comprehension with specific AI levels.
Because the AI ranges from 0 to 1, you might assume that an AI of 0.5 yields 50% comprehension. In fact, a person can understand more than 95% of sentences in this case. Meaningful levels of privacy are only achieved once you get down to AI ratings of less than 0.2.
In order to make things clearer, the Articulation Index range has been divided into subjective speech privacy ranges, which also highlight the need for low AI levels. From 1.0 down to 0.3, you actually have no privacy at all, with comprehension above 80%. From 0.3 to 0.2, you have ‘Marginal’ privacy, with comprehension between 50 and 80%. Such a rating represents only a slight reduction in intelligibility. Conversations are still clear and distracting. From 0.2 to 0.05, you have ‘Normal’ privacy, with 10 to 50% comprehension. Speech is less intelligible and less distracting. Finally, from 0.05 to 0, you have ‘Confidential’ levels of privacy, where less than 10% of a conversation is understood.
While other methods for quantifying speech intelligibility have been developed over the years, the Articulation Index remains in wide use. Next week, I’ll discuss (okay...criticize) a more recent metric: the Privacy Index.