File: tones.doc, updated 8/8/95

			       FFMTIMIT

	      Acoustic-Phonetic Continuous Speech Corpus
		   Far Field Microphone Recordings

                     Training and Test Data
                     NIST Speech Disc 21-1.1


At the beginning of each data collection day, a calibration tone was
recorded using the same B&K microphone and equipment used to collect
the speech data.  The calibration tones were produced with the use of
a B&K type 4230 acoustic calibrator which produces a 94 dB calibration
tone with respect to a reference sound pressure of 0.002 microbars, or
2 * .0001 * (N/m*m), an internationally accepted reference pressure
for 0 dB sound pressure level.

These calibration tones may prove to be useful in making inferences
about vocal effort or background noise level.

NIST's analyses of data from the calibration tones are conventionally
referred to a 0 dB value of a sample amplitude of unity.  With respect
to this reference level, typical calibration tones are approximately 82
dB, corresponding to an rms signal amplitude of ~12750.

The approximately 12 dB difference between the 94 dB sound pressure
level, and the 82 dB level, referenced to unity sample amplitude, may
be useful in making inferences about speech level and background noise
level. 

For example, NIST's measurements of the broadband (background) noise
level are typically about 39 dB relative to unity sample amplitude.
These may correspond to a sound level of 51 dB.  Similarly the levels
corresponding to the 95th percentile of the histograms of the speech
power, with the modeled noise distribution subtracted out, are
typically about 54 dB, corresponding to sound levels of approximately
66 dB.  Broadband signal-to-noise levels are about 15 dB for the B&K
data.

When these data are processed with an A-weighting bandpass filter, the
amplitude of the calibration tones is unaffected.  However, both the
background noise levels and speech levels are reduced.  In particular,
the A-weighted background noise levels are substantially reduced, to
approximately 9 dB relative to unity sample amplitude suggesting an
A-weighted sound level of approximately 21 dB.  The levels
corresponding to the 95th percentile of the histograms of the
A-weighted speech power, with the modeled noise distribution
subtracted out, are approximately 50 dB, corresponding to A-weighted
sound levels of 62 dB.  A-weighted signal-to-noise levels are thus
approximately 41 dB.

The calibration tones are located in the "/tones" directory.  The
files are named by the date recorded (i.e., yymmdd.wav).