A subjective performance study of a sinusoidal speech coding model

Recently, there has been an increasing interest in developing wireline quality speech coders at rates of 4 kbps and below. Good quality at low bit rates is often claimed to be achieved with spectral domain coders which are based on a sinusoidal representation of the speech signal. However, based for example on the recent efforts to standardize a 4 kbps speech coder in International Telecommunications Union (ITU) the current technology has proved to be immature to meet the wireline quality requirements. In this paper, the quality of a speech coder implementation which is based on a commonly used sinusoidal model is verified by a listening test. Three coder versions are tested to evaluate the performance of the underlying speech coding model itself as well as the performance of the widely used parameter models which make sinusoidal coding amenable for low bit rates. Based on the test results, the quality of the implemented sinusoidal speech coding model is not quite comparable to that of the ITU Recommendation G.726 (ADPCM) at 32 kbps, which is commonly used as a reference for wireline quality. Additionally, it was observed that the implemented phase model degrades clearly the performance of the underlying sinusoidal coding model.