Robust Speech Rate Estimation for Spontaneous Speech

Abstract

In this paper, we propose a direct method for speech rate estimation from acoustic features without requiring any automatic speech transcription. We compare various spectral and temporal signal analysis and smoothing strategies to better characterize the underlying syllable structure to derive speech rate. The proposed algorithm extends the methods of spectral subband correlation by including temporal correlation and the use of prominent spectral subbands for improving the signal correlation essential for syllable detection. Furthermore, to address some of the practical robustness issues in previously proposed methods, we introduce some novel components into the algorithm such as the use of pitch confidence for filtering spurious syllable envelope peaks, magnifying window for tackling neighboring syllable smearing, and relative peak measure thresholds for pseudo peak rejection. We also describe an automated approach for learning algorithm parameters from data, and find the optimal settings through Monte Carlo simulations and parameter sensitivity analysis. Final experimental evaluations are conducted based on a portion of the Switchboard corpus for which manual phonetic segmentation information, and published results for direct comparison are available. The results show a correlation coefficient of 0.745 with respect to the ground truth based on manual segmentation. This result is about a 17% improvement compared to the current best single estimator and a 11% improvement over the multiestimator evaluated on the same Switchboard database.

Use of syllable nuclei locations to improve ASR

“…This method can produce spurious peaks in non-speech and unvoiced regions, so a pitch detector is applied to the waveform and all peaks corresponding to unvoiced and non-speech segments are removed. In [4], it was reported that this method correctly placed nuclei in 80.6% of the syllables in a hand transcribed test set. In [3,4], peaks that fall below a minimum threshold are rejected and the result is a binary feature.…”

“…In [4], it was reported that this method correctly placed nuclei in 80.6% of the syllables in a hand transcribed test set. In [3,4], peaks that fall below a minimum threshold are rejected and the result is a binary feature. For our experiments we do not make a hard decision, instead we retain all the maxima points and use the actual height value as a feature.…”

“…Before this DBN is presented, the feature extraction process is described. This process was given by Wang in [3,4]. First, a 19 band filter is applied to the waveform, and the 5 bands with the most energy are selected.…”

“…The second goal is to use long time scale features to influence duration and segmentation directly, without having to "pass through" a sub-phone state variable. The particular acoustic queues used are estimates of syllable nuclei locations derived from a spectral correlation envelope [2,3,4]. A dynamic Bayesian network (DBN) is used to integrate a state variable that counts syllable nuclei with a traditional recognizer (that uses a WIP).…”

“…Syllable nuclei estimates via a spectral correlation measure were first used to estimate speaking rate (one of the 3 measures in mrate) [2]. This idea was expanded on by Wang to include temporal correlation and a number of other improvements [3,4], and this is the method employed in this work. Wang used this detection method in [12] to create speaking rate and syllable length features for automatic speech prominence detection.…”