This paper presents a new approach to speech synthesis in which a set of cross-word decision-tree state-clustered context-dependent hidden Markov models are used to define a set of subphone units to be used in a concatenation synthesizer. The models, trees, waveform segments and other parameters representing each clustered state are obtained completely automatically through training on a 1 hour single-speaker continuous-speech database. During synthesis the required utterance, specified as a string of words of known phonetic pronounciation, is generated as a sequence of these clustered states using a TD-PSOLA waveform concatenation synthesizer. The system produces speech which, though in a monotone, is both natural sounding and highly intelligible. A Modified Rhyme Test conducted to measure segmental intelligibility yielded a 50% error rate. The speech produced by the system mimics the voice of the speaker used to record the training database. The system can be retrained on...

Corpus-based approaches to speech synthesis have been advocated to overcome the limitations of concatenative synthesis from a xed acoustic unit inventory. The frequency of unit concatenations in, e.g., diphone synthesis has been argued to contribute to the perceived lack of naturalness of synthetic speech. The key idea of corpus-based synthesis, or unit selection, is to use an entire speech corpus as the acoustic inventory and to select at run-time from this corpus the longest available strings of phonetic segments that match a sequence of target speech sounds in the utterance to be synthesized, thereby minimizing the number of concatenations and reducing the need for signal processing. This paper reviews the assumptions underlying this synthesis strategy and the dierent approaches to unit selection, as well as the major challenges encountered by corpus-based methods. One of the biggest problems to date is the relative weighting of acoustic distance measures. We further argue agains...

Concatenative "selection-based" synthesis from large databases has emerged as a viable framework for TTS waveform generation. Unit selection algorithms attempt to predict the appropriateness of a particular database speech segment using only linguistic features output by text analysis and prosody prediction components of a synthesizer. All of these algorithms have in common a training or "learning" phase in which parameters are trained to select appropriate waveform segments for a given feature vector input. One approach to this step is to partition available data into clusters that can be indexed by linguistic features available at runtime. This method relies critically on two important principles: discrimination of fine phonetic details using a perceptually-motivated distance measure in training and generalization to unseen cases in selection. In this paper, we describe efforts to systematically investigate and improve these parts of the process.

This paper proposes a two-module TTS structure, which bypasses the prosody model that predicts numerical prosodic parameters for synthetic speech. Instead, many instances of each basic unit from a large speech corpus are classified into categories by a CART, in which the expectation of the weighted sum of square regression error of prosodic features is used as splitting criterion. Better prosody is achieved by keeping slender diversity in prosodic features of instances belong to the same class. A multi-tier non-uniform unit selection method is presented. It makes the best decision on unit selection by minimizing the concatenated cost of a whole utterance. Since the largest available and suitable units are selected for concatenating, distortion caused by mismatches at concatenated points is minimized. Very natural and fluent speech is synthesized, according to informal listening test.