Indoor Acoustics: Restaurants and noise

I've recently begun following a blog called Cross-Spectrum/Acoustics - Sound, Vibration, and a Little Noise.  It has quite an interesting collection of links to articles on various topics on acoustics and noise, as the title implies, as well as good original content by the author.

On to the topic of this post.  One of the perennial challenges for forensic audio cleanup is to reduce the amount of interference on recordings made in bars, pubs, and restaurants.  In recent years, say the past ten, the amount of noise found in restaurants in the USA has risen dramatically thanks to a trend in restaurant design, which has resulted in an increase in the number of cases requiring forensic restoration and enhancement. Cross-Spectrum has a post with links to recent articles on the restaurant noise phenomenon.

"Restaurant noise" is a little too broad of a term in itself to describe the problem from a forensic standpoint, however.  Consider the different types of noises that may be found there:

• music (instrumental and/or vocal; live or pre-recorded)
• talking (speech; the cocktail party problem of too many voices blending together into babble)
• dropped flatware (knives, forks)
• clinking glassware
• shouts (e.g. to get attention of wait staff)
• resonance and reverberation
• hum (fluorescent lighting)

Forensic audio practitioners will recognize lots of things about this list, for instance, impulsive noises (flatware, glassware, and shouts), broadband noise (talking), and harmonic noises (instrumental noise, hum and resonance).  Each of these types of noises is different and may require different treatment; however, one of these noises rises above the others in difficulty and that is talking (speech).  

Filtering out noise generally requires recognizing the noise signal as being somehow different from the desired signal, which is usually speech, in some domain we can measure it in (i.e. time, frequency, amplitude, or space) and then removing the noise using a filter implemented in the corresponding domain.  That process falls apart when the noise is speech itself - how do you tell the difference between interfering speech and desired speech on a mono, omni-directional recording made in an uncontrolled environment?  Therein lies the problem.  There are practical solutions, as well as impractical ones for that matter, but the best ones require prior-preparation.  Unfortunately, I'll have to leave this filtering discussion at this point as going in much deeper could lead to a book-length article!  (If you are interested in going in deeper, feel free to ask away in a comment or contact me directly using the email link found near the top-right of the page.)

As you can hopefully now appreciate, our job would be so much easier without this recent trend in designing restaurants, bars, and pubs to strengthen the cocktail party problem, hence my interest in the Cross-Spectrum posting.