Cracking sounds emitted by coffee beans during the roasting process were recorded and analyzed to investigate the potential of using the sounds as the basis for an automated roast monitoring technique. Three parameters were found that could be exploited. Near the end of the roasting process, sounds known as “first crack” exhibit a higher acoustic amplitude than sounds emitted later, known as “second crack.” First crack emits more low frequency energy than second crack. Finally, the rate of cracks appearing in the second crack chorus is higher than the rate in the first crack chorus.
Coffee is the world's most widely traded tropical agricultural commodity, according to the International Coffee Organization, a global intergovernmental trade group. In the 2011/12 season, 134.4 × 106 bags of coffee (60 kg each) were exported by countries that produce coffee beans, worth an estimated $30.1 × 109.1 By one estimate this yields 1.5 × 109 servings of coffee consumed every day, worldwide.2 Green coffee beans must be roasted before they are used in all forms of the coffee beverage. Roasting the beans is accomplished using a variety of heating methods and at a variety of scales, ranging from mass-market industrial roasters running continuously (processing as much as 5000 kg/h) to the single batch home roaster processing a couple of batches a week (<1 kg/h). To place the economic impact of global coffee roasting in perspective, the cost of energy required to roast the world's yearly supply is about $1 × 109 (calculated using the average 2011 consumer cost of electricity in the US, $0.1/kW-h), hence both economic and quality optimization is of interest.
Controlling the roast time and temperature profile results in a range of roast levels from light to dark, greatly affecting the style, flavor, and aroma of the resulting coffee beverage. Terminating the roasting process at just the right time allows the roaster (most often, a human operator) to achieve the desired darkness of the roast and its accompanying flavor profile, and hence is one of the key roast parameters. Several metrics can be monitored (time, color, aroma, bean volume, bean temperature), using process measurement instrumentation or by the person conducting the roast, to indicate the degree of roasting and ultimately to determine when to terminate the roast.3
The roasting process can also be monitored by ear, by listening for events known collectively as “first crack” and “second crack” (described more completely below), which also signify the progression of the roast. Depending on the type of roasting machine, these sounds can usually be heard by the unaided ear and are perhaps the most important way to monitor the roast for small-batch and artisanal roasting, often combined with the visual cue of bean color and the olfactory cue of aroma. For automated roast monitoring, a number of techniques have been studied, such as monitoring bean volume and porosity,4 content of the exhaust gases,5 and color of the beans,6 among others. Despite the widespread practice of monitoring the roasting process audibly, in the home, commercial artisanal, and mass-market industrial roasting venues, the author has found no previous quantitative description of the sounds produced during coffee roasting, and no discussion of an automated acoustic monitoring technique.
Briefly, the sounds known as first crack and second crack are the result of pyrolytic processes that occur within the bean during roasting. First crack occurs at an internal bean temperature of approximately 200 °C and coincides with the release of steam and gases. Each bean experiences a first crack, and hence as a group of beans is roasting, first one or two cracks are heard and then more and more occur forming a chorus of first cracks. After about two minutes, first crack ceases and there is an acoustically inactive time. Second crack occurs when the temperature reaches about 230 °C and additional gas is emitted along with increased fracturing of bean material. Again a few beans begin second crack. Over about the next two minutes, a chorus builds, peaks, and declines and then second crack stops. The beans can ignite and burn beyond this phase and hence roasting rarely continues beyond the end of second crack.7,8
The sounds of first crack are qualitatively similar to the sound of popcorn popping while second crack sounds more like the breakfast cereal Rice Krispies® in milk. Additional qualitative audible differences between first and second crack are: first crack is louder, first crack is lower in frequency, and individual second cracks occur more frequently within the chorus than first cracks. The purpose of the present work is to quantify these effects as a preliminary step toward the development of an automated acoustical roast monitoring technique.
II. Description of the measurements
A consumer-grade, 0.45-kg-capacity, drum-based coffee roaster with an electrical heating element (1.6 kW) was used to roast a 0.23 kg batch of green coffee beans through the end of second crack. The beans were a typical blend of Arabica and Robusta beans marketed as an espresso blend. A Roland R-26 portable digital audio recorder was used to record the sounds emitted during the roasting process, specifically the first and second crack choruses previously described, including the sounds of the roasting machine itself. The roaster was operated outdoors on a flat concrete surface to eliminate potential reverberation encountered in an enclosed space. The recorder was mounted on a short tripod (15 cm height) that was placed on the concrete surface, at a distance of 0.35 m from the roaster. The recorder's automatic gain function was disabled and one of the R-26's built-in microphones was used. A free-field calibration of the recording system was performed using a substitution method in the fully anechoic chamber at The University of Texas at Austin, so that absolute acoustic pressure could be determined from the data. The frequency response of the recording system was found to be flat within ±2 dB from 20 Hz to 40 kHz. After recording, the data was transferred to a laptop computer for analysis. Minute-long segments of the recordings are presented in Mm. 1 and Mm. 2, for first and second crack, respectively. Note that both files were amplified by 18.2 dB for optimum playback but relative gain was preserved.
Initial analysis indicated that cracking events could be detected automatically using thresholding above +23 mPa and below −23 mPa. The analysis presented in this section was conducted on all the cracks found above this amplitude threshold, which resulted in detection of 62 cracking events in the first crack chorus and 241 events in the second crack chorus.
To quantify the qualitative assessment that first crack is louder than second crack, the peak acoustic pressure (scaled to 1 m by spherical spreading) of individual cracking events was analyzed and histograms were formed for the events of first and second crack. These crack amplitude distributions are shown in Fig. 1, where it can be seen that the maximum amplitudes of first crack and second crack are 63 mPa and 55 mPa, respectively, and that there are a larger number of higher amplitude events in first crack. Hence, a simple peak finding process could be used in an acoustic process control system to automatically differentiate between first crack and second crack using peak acoustic pressure.
To quantify the second qualitative assessment, that first crack events are lower in frequency than second crack events, averaged acoustic pressure spectra were calculated. Ten individual cracks were taken at random from within both first and second crack choruses. These events included both the cracking sounds and the noise due to the roaster. The pressure signatures were detrended, then 512-point, Hann windowed fast Fourier transforms (FFTs) were applied and averaged. These averaged acoustic pressure spectra are shown in Fig. 2. First crack contains more low frequency energy, with a spectral peak at about 800 Hz. Second crack exhibits a flatter spectrum that is lower in amplitude than first crack up through about 10 kHz, and has a spectral peak at about 15 kHz. The background noise emitted by the roaster, including the sounds of the beans rotating in the drum, and the sound of a fan circulating heated air, but absent of any cracking events was analyzed in the same way, except using 4096-point FFTs, to better resolve the tonal components of the roaster noise. Ten segments of crack-free noise were used to determine the average noise spectra due to all other aspects of the roasting process. Figure 2 shows broadband noise due to beans rotating in the drum and the broadband component of fan noise, which can be seen along with spectral lines due to the rotating machinery. To further emphasize the mean characteristics of the broadband noise, a smoothed spectrum is also shown in the thick black curve. In all cases, the noise level is sufficiently below the level of the cracking events. These results indicate that the mean spectral content of just ten individual cracks, including roasting process noise, can be used to automatically differentiate between first crack and second crack acoustically, using a relatively low-resolution FFT, even when no special care is taken to reduce or exclude the roasting process noise.
Finally, the rate of emission of individual cracks was analyzed for both first and second crack choruses as a function of roast time and is shown in Fig. 3. First crack progresses from just before 400 s within the roast and ends at about 600 s. Second crack begins at about 620 s and ends at about 730 s. The qualitative assessment, that second crack events occur more frequently than first crack events is verified. Figure 3 shows that first crack has a peak rate of about 100 cracks per minute, while second crack has a peak rate of over 500 cracks per minute. These results also indicate that an automated technique could be used to differentiate between first crack and second crack by rate, providing a third metric for the automatic acoustic monitoring of the roast.
Sounds emitted during the coffee roasting process were measured and analyzed, including the sounds of first crack and second crack, and the background noise produced by the rotating drum and by the circulating fan. Three acoustical characteristics of the process were found that could be used to form an automated acoustical roast monitoring technique: first crack is louder than second crack (by 15% in peak acoustic pressure), first crack is significantly lower in frequency than second crack (by a factor of nearly 19), and second crack events proceed at a higher rate (by a factor of about 5) than first crack events. Other roasting noise does not impact the use of these signals. These factors were quantified for an espresso blend and one particular roasting machine. Future work should include analysis of all the various types of green coffee beans of interest and different roasting processes and machines.
The author acknowledges support of The University of Texas at Austin Department of Mechanical Engineering and the Fluor Centennial Teaching Fellowship #2.