Revel Salon loudspeaker A Visit, Continued 2
Voecks: The real breakthrough in the design of the Salon was getting a speaker measurement that correlated with double-blind listener preferences. We divided our measurements into three groups.
First was the direct sound. We measure direct sound at the listening window. This yields a curve that gives loudspeaker measurements a bad name, because it can look good with a lousy speaker. We do an average of seven measurements. That cancels out local interference that doesn't mean anything in itself. Other designers may smooth the curve so it can be legible, but that throws out important data. By taking the average over a small area, we end up with a smooth, interpretable curve. This is because the artifacts average out, without having to discard data that turn out to be important.
Second, the early reflections from nearby surfaces are measured after a slight delay—20ms. This is so early that the ear-brain system can't sort it out. The performance of the speaker 60-70 degrees off-axis needs to be included in this early-reflections measurement because it contributes to early reflections. The off-axis sounds are just as important as the on-axis. Two speakers that look the same when you measure the direct sound can look wildly different off-axis.
Third, the off-axis speaker performance contributes most to the reverberant sound. This correlates with speaker timbre, one of the most important factors in listener preferences during double-blind tests. We take 72 measurements in the "4Pi" anechoic chamber by rotating the loudspeaker on a table and varying microphone height. The listening tests conducted in the MLL enabled Floyd and the group to determine how to use these 72 measurements and weight them to generate a single curve. This generates the Calculated Perceived Response, which we call our "target." The ideal, as far as listeners go, is not a flat line. If [a loudspeaker is] set to be flat, it is perceived as overly bright. Listeners prefer the "target" to be sloping down, and report that as "neutral."
Greenhill: As the Salon's designer, you used the double-blind listening panel's preferences to confirm your design choices. Was this panel trained to a standard before these tests? If so, this would be the first use of a trained and "certified" double-blind listening panel to help a designer!
Voecks: Yes, all panel members are trained first using a computer-aided instruction course before they participate in listening tests in our MLL (footnote 2). Sean Olive, Harman's manager of subjective evaluation, explains that "the current threshold or criteria in selecting listeners on our 'expert' or 'trained' panel is that they must achieve at least 95% correct in the 6dB-resonance identification task and 85% correct in the 3dB task. Both tasks involve correctly identifying the center frequency of resonant peaks and dips of added resonances that are 2.5 octaves wide. Generally, members who meet these criteria do so after one to three training sessions that last about 30 minutes each. We have found that after five training sessions, about 20% of motivated people with normal hearing cannot meet this target and are not accepted on the panel. In addition, we do not select listeners with more than -15dB loss at any audiometric frequency, since Toole found that listeners ! with hearing losses greater than this are less consistent in their opinions compared to normal-hearing listeners."
Footnote 2: See S.E. Olive, "A Method for Training Listeners and Selecting Program Material for Listening Tests," presented at the 97th Convention of the Audio Engineering Society (November 1994). Preprint 3893 (H-5).