Scott A. Amman - 73121.351@compuserve.com
Ford Motor Co.
3003 Prairie
Royal Oak, MI 48073
Popular version of paper 2aEA6
Presented Tuesday morning, 14 May 1996
Acoustical Society of America, Indianapolis, Indiana
Embargoed until 14 May 1996
Basically, what we're trying to do here is to apply techniques used in speech modeling and synthesis to the production of automotive sounds. You may ask why anybody would want to synthesize automotive sounds. We (Ford Motor Co.) have a Sound Quality Lab where we try to find out what customers^R likes and dislikes are when it comes to the total acoustic environment in the car. If we have the capability to artificially create sounds (such as that from a running engine) we can present these to customers to determine what they desire in their vehicles. This paper deals with engine sounds. In the paper we draw parallels between the engine noise production mechanisms and those of speech. Many aspects of engine noise generation parallel that of human speech production. The noise from an internal combustion engine is typically experienced from the passenger compartment of the vehicle. The engine noise is a complex signal produced by a number of sources in the engine compartment. These sources can be broadly classified as either deterministic (predictable) or stochastic (random) in nature. An analogy can be made between these two classes, and voiced and unvoiced speech. Unlike speech, however; both deterministic and stochastic components occur simultaneously for engine noise.
Deterministic Component
A large part of the deterministic component is due to the combustion of fuel in the engine cylinders. This produces pressure pulses not unlike those produced by the glottis in voiced speech. In this paper we will discuss some of the components in speech production and compare these to that of sound production of a multicylinder internal combustion engine. The engine produces these pulses at a rate proportional to engine speed. For a four stroke engine, fuel is ignited once every two engine revolutions for each cylinder. Thus, an eight cylinder engine produces four pressure pulses per crankshaft revolution. Pressure pulse frequency is like that found in pitch extraction for speech in which the frequency of the glottal pulses are determined. As with speech, the source manifests itself spectrally as a frequency component at pulse frequency and its harmonics. Other deterministic sources include engine accessories such as the alternator, AC compressor, engine cooling fan (which also produces stochastic noise due to turbulent air flow), etc. These sources produce sinusoidal components which are highly correlated to the engine combustion since they are driven by the engine crankshaft. Depending on pulley ratios, these components do not necessarily represent multiples of the pressure pulse frequency . This is unlike voiced speech production in which only the pressure pulse frequency and its harmonics exist.
Stochastic Component
There is also a stochastic sound component produced by the engine. This is primarily due to turbulent air flow from the air intake and exhaust systems. Turbulent air flow can also occur from accessories such as the engine cooling and alternator fans. This is like the production of non-voiced sounds in speech in which colored noise is generated by turbulent air flow over the tongue and lips.
If only the airborne component of the engine noise (coming from the engine compartment) is considered, both deterministic and stochastic components are subjected to the same transfer function before reaching the passenger compartment. For voiced speech this transfer function would be analogous to the transfer function of the vocal tract which spectrally shapes the glottal excitation. Unlike unvoiced speech production, the stochastic component of the engine noise is subjected to the same transfer function as the deterministic component. This transfer function is dependent on the materials and structure used to isolate the passenger from the engine. As mentioned previously, speech production is generally classified as voiced or unvoiced. Therefore, segmentation of the speech signal is required before traditional synthesis techniques can be implemented. This is not a concern for engine noise synthesis since both components occur simultaneously.
The preceding analysis gives us reason to believe that there may be speech synthesis techniques which can be applied to the problem of engine noise synthesis. We first briefly examine regular or periodic pulse excitation (pulses occur at a constant rate and height, and are spectrally shaped by a filter) and discuss some of the pitfalls of this approach for engine noise generation. Next, we introduce a new suboptimal multipulse (pulses occur at varying intervals and heights and are spectrally shaped by a filter) technique for the purpose of engine noise production. In this approach the deterministic components, as measured in the vehicle passenger compartment, are extracted from the signal. The remaining stochastic component is then synthesized by a multipulse excited time series model (the time series model is actually just a filter). Together, the deterministic and stochastic components are added to generate the synthesized sound. The resulting synthetic sound actually requires much less storage (2% of the original) space. This is an added advantage when trying to store or transmit long sound files.