Improved Epoch Based Prosody Modification by Zero Frequency Filtering of Gabor Filtered Telephonic Speech
M. Rama Rajeswari, D. Govind, Suryakanth V. Gangashetty, Akhilesh Kumar Dubey · 2023
Prosodic features are supra-segmental features which span over longer segments in speech signals. Modification of prosodic parameters is essential for some of the applications such as text to speech synthesis, emotion conversion, speech enhancement and so on. Prosody modification is achieved by the manipulation of pitch contours and duration of the given speech. Estimating the analysis pitch marks is one of the crucial steps in prosody modification of speech due to vocaltract system interaction during the phonation. The presence of noises, channel effects and other degradations make the epoch estimation even harder. Therefore, the context of the work presented in this paper is targeted towards epoch estimation and subsequent prosody modification from speech signals in telephone quality. In the case of telephonic speech signals, due to channel bandwidth (300 Hz-3.4 kHz), fundamental frequency components are significantly attenuated and therefore estimating epochs in such cases are challenging. In the present work, fast and efficient zero frequency filtering (ZFF) is used as the method for hypothesizing the epochs in telephonic speech. Due to attenuation of the low frequency components (less than 300 Hz), ZFF provides poor estimation of epochs in the case of telephonic speech signals. The strength of impulse-like discontinuities has to be enhanced for improved epoch estimation from telephonic speech using ZFF method. Further, a multiscale product of the Gabor filtered speech with given speech is proposed as a preprocessing step for improved epoch estimation using ZFF method. The improved epoch estimation performance of the proposed preprocessing method is confirmed by evaluation on the simulated telephonic version of the CMU-Arctic database where the ground truth reference epochs are available in simultaneous electro-glottogram recordings. Based on subjective evaluations, the proposed epoch estimation provided perceptually better prosody modified telephonic speech.