Citation Context

...d ratio scores. The most prevalent methods include H-norm [14], Z-norm [15], and T-norm [16]. Examples of the model-domain compensation methods include the speaker-independent variance transformation =-=[17]-=-, and the transformation for synthesizing supplementary speaker models for other channel types from multichannel training data [18]. Additionally, channel mismatch has also been dealt with by using mo...

Citation Context

...the problem. The issue of acoustic mismatches can be tackeled at different levels: speech features can be extracted that are less sensitive to channel effects than the traditional cepstrum (see e.g., =-=[2, 13]-=-) the effect of mismatches can be reduced via cepstral mean/bias removal (see e.g., [3-6]), the speaker models can be transformed to compensate for the mismatches, rescoring techniques can be used to ...

Citation Context

...m achieved over 94% recognition rate on the KING Speech Corpus database [8]. Several authors have proposed features calculated starting from the ones described by Reynolds and Rose. In particular, in =-=[9]-=- a transform was applied to the Mel cepstral features in order to compensate the noise components of the audio channel. From the transformed domain, the so-called formants features were then calculate...

Citation Context

...ers and measured SI accuracy of approximately 60% (40% lower compared to TIMIT). In [6], the authors considered the problem whereby a SI system is trained with clean speech but tested with speech acquired over lossy, packet channels, such as cellular and VoIP. Accuracy levels for SI using the YOHO speech corpus were reported as 30–60% depending on the packet loss rates. In addition to evaluating SI performance under various channel conditions, many methods have also been proposed to compensate or equalize speech signals in order to improve SI accuracy with varying degrees of success [2], [7], [8], [9], [10]. Likewise, in SV applications many methods have also been proposed including factor analysis [11]. In the majority of this work, the channels under consideration have been various microphone types, telephone and mobile channels, or VoIP. In this paper, we consider the impact of acoustic room reverberation on SI (not SV). We specifically investigate the problem where we have access to clean training signals but have acquired the test signals in a reverberant environment. Such a mismatch between training and test signals can easily occur in audio surveillance applications where highe...

Citation Context

...coefficients were considered to construct a feature vector. IV. SPEAKER MODELLING USING GMM AND HMM Gaussian mixture model is the classifier commonly used in speaker identification/ verification [1], =-=[14]-=- and [15]. This classifier is able to approximate the distribution of the acoustic classes representing broad phonetic events occurring in speech production (e.g., during the production of vowels, nas...

Citation Context

...be separated through cepstral decomposition (i.e., spectral slope vs. spectral detail, respectively). While F0, vocal tract length and spectral slope are known to be good discriminator among speakers =-=[14]-=-, additional acoustic cues from the learner’s voice could be captured and preserved before the morphing stage. As an example, jitter and shimmer (cycle-tocycle variations in fundamental frequency and ...

Citation Context

...be separated throughscepstral decomposition (i.e., spectral slope vs. spectral detail,srespectively). While F0, vocal tract length and spectral slope aresknown to be good discriminator among speakers =-=[23]-=-, additionalsacoustic cues from the learner’s voice could be captured andspreserved before the morphing stage.sAs an example, jitter andsshimmer (cycle-to-cycle variations insand amplitude,srespective...