Citation Context

...g produced/uttered by human-beings. For example, it is known that the prosody and acoustics of a word spoken in a sentence significantly differs from that of spoken in isolation. The work done in [3] =-=[4]-=- [5] [6] [7] [8] suggests that a similar analogy of prosodic and acoustic difference exists for sentences spoken in isolation versus sentences spoken in paragraphs and similarly for paragraphs too. So...

Citation Context

...honetic transcriptions for the purpose of learning probabilistic phonological rules [158]. Closer to our purpose, Prahallad altered the architecture of a speech decoder at the level of phone topology =-=[126]-=-. The topologies were either a) linear 5-state sequential with no skip states, b) fully connected 5-state networks, or c) linear 5-state sequential with fully connecting forward jumps. The states are ...

Citation Context

...t and long vowels could improve our models. We decoded the speech data with the previous English acoustic model. We then aligned the speech to these phonetic transcripts using an EHMM alignment tool (=-=Prahallad et al., 2006-=-). We determined the duration of different vowels and clustered them into two groups based on the duration. We then labeled these vowel clusters as being two different vowels when training the synthet...

Citation Context

...r labelling speech data is manual labelling by linguistic experts. This task is both time consuming and complicated, therefore automatic methods and algorithms have been developed over the last years =-=[1, 2, 3]-=-. Automatic labelling methods are still error-prone, so that they are often followed by a stage of manual correction, which can be directed by a confidence information that indicates which labels to v...

Citation Context

... HMM/Multi-Layer Perceptron (MLP) acoustic model combining posterior phone probabilities generated by several MLP’s trained on distinct input features [7,8]. The MLP network weights were re– estimated to adapt the models to the speaker; Hidden Markov Model Toolkit: (HTK) [9], using unsupervised speaker-adapted, context-independent Hidden Markov Models (HMM). The models were adapted based on initial segmentations generated by the MAF–DTW tool. The models have ergodical left–right topology, with 5 states each (3 emitting states); eHMM: phonetic alignment tool oriented for speech synthesis tasks [10] , developed in Carnegie Mellon University and distributed together with a set tools for building voices for Festival, called Festvox 2.1 [11]. The adopted model topology is the same as described for HTK; eHMM was also used doing acoustic model adaptation to the speaker. In Table 1 we present the overall performance of each segmentation tool. From this table, it can be seen that the MAF–DTW is the tool with the worst performance in terms of Absolute Mean Error (AME): 41ms. This value is almost twice as much as the second worst result (eHMM). This was already expected, as DTW algorithms are usu...

Citation Context

..., “audio-book” but isdifferent from the TATS corpus. We extracted the pause duration from natural speech in our corpora. To do that, we force-aligned the speech and transcriptions using an EHMM tool =-=[16]-=- that allows for short silences to be inserted during the alignment. We used these alignments to find out the length of these inserted silences. We ignored all inserted pauses that were less than 80ms...

Citation Context

...tion modeling in isolation, and note that it is an important problem in its own right (as part of the study of phonology) as well as necessary for other tasks (such as conversational speech synthesis =-=[11]-=-). 2. Feature-based pronunciation modeling The pronunciation models we consider are based on ideas from autosegmental [3] and articulatory [1] phonology. In such a model, the typical single sequence o...

Citation Context

...by others. 2.1. Training The first stage, which is not technically part of the CLUSTERGEN synthesizer is to label the database using an HMM labeler. For the results presented here, we have used EHMM, =-=[11]-=-, which is included within the latest FestVox release. It uses Baum Welch from a flat start to train context independent HMM models, which it then uses to force align the phonemes generated from the t...

Citation Context

...speech analysis respectively. In turn they produce phone sequences and signal features, fundamental frequency, mel-cepstral coefficients (MCEP) and energy. Phone sequence and MCEPs are passed to EHHM =-=[10]-=- for labeling the speech signal with respect to phone sequence of the utterance. EHMM would produce the labels with phones and it’s time stamps in the speech signal. Rest of the procedure is broken in...