Phase-corrected RASTA for automatic speech recognition over the phone

Effectiveness of phase-corrected rasta for continuous speech recognition

“…The present set o f experiments indicates that a successful CN m ethod should not introduce any phase distortion if CD-HM M s are used and the training data is not sufficient to m odel the left context dependency for all relevant contexts. This result is in good agreem ent with the conclusions in [4,5,6].…”

“…However, for more complex acoustic models the performance differences become in significant and pcR performs as w ell as CMS. The results shown in Figure 2 are in good agreem ent w ith the re sults we reported earlier in the context o f a connected digit recog nition task [5,6]. In that case we also found that pcR was capable o f outperforming clR, that pcR w as preferred over N CN and that pcR and CMS performance showed no significant difference.…”

“…We used clR w ith integration factor -0.94 [2]. For pcR we used the same integration factor in combination w ith a phasecorrection filter [4,5,6]. During the time-reversal operations re quired for the phase-correction we used the whole utterance [6].…”

“…[1,2,3]). Recently, a new, extended ver sion o f the classical RASTA filtering technique was proposed and tested in the context o f connected digit recognition over the tele phone [4,5,6]. The results o f these connected digit string experi m ents showed that the recognition performance o f phase-correct ed RASTA (pcR) is equivalent to the performance o f cepstrum m ean subtraction (CMS).…”

SPECTRA: A SPEech proCessing plaTform as smaRtphone Application

“…[4] suggests to extract from raw audio samples the Mel Frequency Cepstral Coefficients (MFCC) and/or the Linear Prediction Cepstral Coefficients (LPCC). Other works, such as [5] employ the RelAtive SpecTral Amplitude (RASTA) coefficients and the Perceptual Linear Prediction (PLP) coefficients. As reported in [2], an increase in classication performance would usually be expected when more features are used.…”