TNS

Temporal Noise Shaping is a special technique to reduce pre-echo artifacts. In traditional block-based transform coding, independent quantization of coefficients results in the coding error being spread in time along the whole block of samples, so if there is a transient - some of the error energy may appear before the transient. TNS applies a preprocessing to the coefficients so that they are quantized in differential manner. A predictor working in frequency domain is used, a prediction error between consecutive coefficients is calculated, independently within each frame. Quantization of frequency-domain prediction error causes the coding error to be shaped so that it resembles the signal in time domain and is more effectively masked.

Tranform coding error without TNS showing pre-echo (left) and with TNS (rigth). Click on both pictures to hear the audio example. Pay attention to the blurring of castanets attack part in the middle of the excerpt.