Modeling Spot Microphone Signals using the Sinusoidal Plus Noise Approach

Abstract

This paper focuses on high-fidelity multichannel audio coding based on an enhanced adaptation of the well-known sinusoidal plus noise model (SNM). Sinusoids cannot be used per se for high-quality audio modeling because they do not represent all the audible information of a recording. The noise part has also to be treated to avoid an artificial sounding resynthesis of the audio signal. Generally, the encoding process needs much higher bitrates for the noise part than the sinusoidal one. Our objective is to encode spot microphone signals using the SNM, by taking advantage of the interchannel similarities to achieve low bitrates. We demonstrate that for a given multichannel audio recording, the noise part for each spot microphone signal (before the mixing stage) can be obtained by using its noise envelope to transform the noise part of just one of the signals (the so-called "reference signal", which is fully encoded).

A Multichannel Sinusoidal Model Applied to Spot Microphone Signals for Immersive Audio

“…In turn, this noise part is synthesized by filtering the residual signal obtained from the reference channel with the time-varying noise envelope of each particular spot signal. This procedure, described in our recent work as noise transplantation [18], is based on the observation that the noise component of the spot signals of the same multichannel recording are very similar when the sinusoidal part has been captured by an appropriate number of sinusoids. For multiple spot signals without any similarities in their content, it is shown in this paper that the same transplantation procedure is valid if the noise component is obtained from a reference signal which is a downmix of the multiple audio recordings.…”

Low Bitrate Coding of Spot Audio Signals for Interactive and Immersive Audio Applications

“…Thus, the objective is to test whether the side signal can be accurately reproduced when using the residual from the reference signal. We note that in our previous work [12], we showed that the proposed noise transplantation approach results in very good quality (around 4.0 grade in DCR tests in most cases) for various music signals, with the number of sinusoids per frame as low as 10. Thus, in this section our objective is to examine the lower limit in bitrates which can be achieved by our system without loss of audio quality below the grade achieved by modeling alone (i.e.…”

“…This is important, since this signal is of highly stochastic nature, and cannot be adequately represented using a small number of parameters (thus, it is highly demanding in bitrates for accurate encoding). We note that modeling this signal with parametric models results in low-quality audio resynthesis; in our previous work [12] we have shown that our noise transplantation method can result in significantly better quality audio modeling compared to parametric models for the residual signal. We obtained subjective scores around 4.0 using as low as 10 sinusoids, which is very important for low bitrate coding.…”

“…For this reason, the noise part of only the reference signal is retained; during the resynthesis of each spot signal, its harmonic part is added to the noise part, which is recreated by using the corresponding noise envelope with the noise residual obtained from the reference channel. This procedure, has been described in our recent work as noise transplantation [12], and is based on the observation that the noise component of the spot signals of the same multichannel recording are very similar when the harmonic part has been captured with an appropriate number of sinusoids. In this paper, we focus on describing the coding stage of the model parameters, and defining the lower limits in terms of bitrate that our proposed system can achieve.…”