This paper presents several findings in comparison of F0 range in Chinese speech by native Chinese speakers (CC), Chinese Speech by Japanese learners (CJ) and Japanese speech by native Japanese speakers (JJ). The purpose of this work as a whole is to investigate possible F0 range difference between Chinese L1 and L2 and examine whether this difference is correlated with their Japanese L1. “Long term distributional” (LTD) measures were adopted, calculated in both sentence and speaker level, and analysis was conducted on a relatively large corpus. Results showed an overall tendency of cross-language differences in the chosen LTD measures suggesting that: 1) F0 range is larger and F0 register higher in native Chinese speech than native Japanese speech; 2) Chinese L2 has a narrower F0 range and lower F0 register than native Chinese speech; 3) Moreover, F0 distribution (kurtosis & skewness) results uncovered a phenomenon that despite the relative consistent trend of narrower F0 range in L2 speakers, it is highly speaker dependent of whether to narrow the upper or lower part of their speaking F0 range.

Formant estimation from vowel segments, is very useful for linguistic purposes. Traditionally, formants are estimated using classical signal processing methods and statistical models. The averaged continuous extraction of formant frequency along the vowel segments is taken as the formant of vowels. New approaches using neural networks to predict vowel formants on an annotated database, where the input is the acoustic features and output is the mean formant frequency value. Recently, the Bilinear Network (BL) and Temporal Attention-Augmented Bilinear Network (TABL) have proven to be very effective on financial time-series analysis task, compared to recurrent networks and convolution networks. Similar to our work, we explored how to extend the structure of BL and learn from TABL to produce better short-term modeling capability for vowel formant estimation. More specifically, we proposed to replace the attention mechanism with sigmoid gate and use a learnable parameter to dynamically integrate the first linear transformation output, thus learning better representation of BL. Experiments on the vowels test set of public VTR corpus showed that our approach significantly surpassed DNN, CNN, BL and achieved slightly better performance than the poweful TABL model in terms of mean absolute error and mean absolute percent error rate on F1, F2, F3 and overall.