Beyond Wideband Telephony - Bandwidth Extension for Super-Wideband Speech

Driven by the market success of high-quality Voice over IP technology, the introduction of wideband telephony with an acoustic bandwidth of at least 7 kHz is meanwhile also foreseen for “traditional” digital voice communication services such as ISDN, DECT, or UMTS. While wideband speech addresses the basic requirement of intelligibility (even for meaningless syllables), the perceived “naturalness” and the experienced “quality” of speech can be further enhanced by providing an even larger acoustic bandwidth. Thus, the next logical step towards true “Hi-Fi Telephony” could be the rendering of “super-wideband” (SWB) speech signals with an acoustic bandwidth of at least 14 kHz. In this contribution, we review previous and current standardization activities with this focus. Moreover, a method for artificial bandwidth extension (BWE) of wideband speech signals towards “super-wideband” is presented and evaluated. It is shown that improved naturalness and speech quality can be attained by a purely receiver-based modification of wideband terminals. Wideband vs. Super-Wideband Speech Typically, wideband (WB) speech is defined by its acoustic frequency range of 0.05 − 7 kHz, whereas superwideband speech provides a roughly doubled bandwidth of, e.g., 0.05−14 kHz. The lower cutoff frequency of 50Hz is usually considered sufficient for a natural reproduction of speech signals. An analysis reveals that on average only about 1.5% of the energy of super-wideband speech signals is located in the 7− 14 kHz extension band (EB). This average is only exceeded in less than 25% of all active frames, which indicates that there must be strong outliers in the EB to SWB energy ratio σ EB/σ 2 SWB. Such outliers are actually found in fricative and plosive speech sounds as illustrated in Fig. 1. For particularly strong outliers, the EB energy is even larger than the WB energy. This is the case for about 6% of all active frames. Here, the largest benefits over WB signals can be expected. In addition to simple energy considerations, there is also evidence that temporal signal characteristics gain perceptual importance with an increasing frequency, cf. [1, 2]. For the EB range of 7−14 kHz, detailed temporal characteristics may be even more important than the exact reproduction of the spectral envelope information. Results are based on approx. 1.6 · 10 active 20ms speech frames, sampled at 32 kHz and low-pass filtered with fc = 14 kHz. 0 2 4 6 8 10 12 14 16