Block-based speech bandwidth extension system with seperated envelope energy ratio estimation

The major issue in extending bandwidth of narrowband speech signal (0–4kHz) is the estimation of high-band portion (4–8 kHz) of spectral envelope. It is found that, apart from the shape of high-band spectral envelope, the relative energy level of the missing high band to the observable low band is also crucial to the system performance. In this paper, the two-fold problem is solved by two different estimation rules. In memoryless bandwidth extension systems, the missing high-band information is estimated from narrowband speech using the current frame only. As the narrowband-to-wideband mapping is a one-to-many problem ([1]), memoryless system is likely to cause hissing and whistling artifacts. Our method treats envelope shape estimation on a block basis. Detected narrowband speech block is either one word or a sequence of words, which is modeled by CDHMM (continuous density hidden Markov model) and mapped to a wideband CDHMM pre-trained by original version of the speech block. High-band energy level, present as normalized energy ratio to observable low-band energy, is estimated on an MMSE rule. Both subjective and objective evaluations show that hissing and whistling artifacts are reduced and the spectrally extended wideband speech (0–8kHz) is pleasant to listen.