Adaptive Postfilter

Description

The adaptive postfilter is a crucial component in many speech codecs, including the GSM Enhanced Full Rate (EFR) codec and the Adaptive Multi-Rate (AMR) codec. Its primary function is to improve the perceptual quality of the reconstructed speech signal by filtering out coding noise and enhancing the formant structure. The postfilter is applied to the output of the short-term synthesis filter, which is the final stage of the speech decoder. By adaptively adjusting its characteristics based on the speech signal, the postfilter can significantly reduce the audible artifacts introduced by the coding process.

In the GSM Enhanced Full Rate codec, the adaptive postfilter is implemented as a cascade of two filters: a formant postfilter and a tilt compensation filter. The formant postfilter is designed to emphasize the formant frequencies of the speech signal, which are the resonant peaks of the vocal tract. It does this by applying a filter that has a frequency response inversely proportional to the short-term spectral envelope. This effectively boosts the formants and attenuates the spectral valleys, where coding noise is more audible. The formant postfilter is typically a short-term filter with coefficients derived from the quantized linear prediction (LP) coefficients of the current frame.

The tilt compensation filter is used to compensate for the spectral tilt introduced by the formant postfilter. The formant postfilter tends to attenuate high frequencies, which can make the speech sound muffled. The tilt compensation filter applies a first-order high-pass characteristic to restore the natural spectral balance. Together, these two filters form the adaptive postfilter, which is applied to the synthesized speech signal. The parameters of the postfilter, such as the formant enhancement factor and the tilt compensation factor, are adapted based on the speech signal characteristics to achieve optimal perceptual quality.

The adaptive postfilter operates on a frame-by-frame basis, typically with a frame length of 20 ms. It uses the LP coefficients from the current frame to compute the filter coefficients. The postfilter is applied in the time domain, and its output is the final reconstructed speech signal. The adaptive nature of the postfilter allows it to adjust its strength depending on the speech signal. For example, in unvoiced segments, the postfilter may be less aggressive to avoid distorting the noise-like signal. In voiced segments, it may be more aggressive to enhance the formants and reduce coding noise.

The adaptive postfilter is also used in other 3GPP codecs, such as AMR and AMR-WB. In AMR, the postfilter is similar in structure but may have different parameter values. The postfilter is an essential part of the decoder and contributes to the overall speech quality. Without it, the reconstructed speech would sound noisy and less natural. The adaptive postfilter is a relatively low-complexity component, but its impact on perceptual quality is significant.

In summary, the adaptive postfilter is a signal processing block that enhances the quality of decoded speech by shaping the spectrum to emphasize formants and compensate for tilt. It is a standard component in 3GPP speech codecs and has been used since the early days of digital cellular communications. Its design and parameters are carefully tuned to provide the best trade-off between noise reduction and speech distortion.

Purpose & Motivation

The adaptive postfilter exists to improve the perceptual quality of speech reconstructed by low-bit-rate codecs. In codecs like GSM EFR and AMR, the speech signal is compressed by removing redundancy, which inevitably introduces quantization noise. This noise is most audible in the spectral valleys between formants, where the speech energy is low. The adaptive postfilter addresses this by attenuating the spectral valleys and boosting the formants, making the speech sound clearer and more natural. It also compensates for the spectral tilt introduced by the formant postfilter, ensuring a balanced frequency response.

Before the adaptive postfilter, early speech codecs often produced speech with noticeable artifacts, such as roughness or muffling. The postfilter was introduced as a post-processing step to mitigate these artifacts without increasing the bit rate. It is a form of perceptual filtering that exploits the masking properties of the human auditory system. By shaping the noise spectrum to be less audible, the postfilter allows the codec to operate at lower bit rates while maintaining acceptable quality.

The adaptive postfilter is particularly important in mobile communications, where bandwidth is limited and speech quality is paramount. It has been a standard feature in 3GPP codecs since the GSM EFR codec and continues to be used in modern codecs like EVS. Its design has evolved over time, with improvements in filter structure and parameter adaptation. The adaptive postfilter is a testament to the importance of perceptual considerations in speech coding.