This paper proposes a technique for adding more prosodic variations to the synthetic speech in HMM-based expressive speech synthesis. We create novel phrase-level F0 context labels from the residual information of F0 features between original and synthetic speech for the training data. Specifically, we classify the difference of average log F0 values between the original and synthetic speech into three classes which have perceptual meanings, i.e., high, neutral, and low of relative pitch at the phrase level. We evaluate both ideal and practical cases using appealing and fairy tale speech recorded under a realistic condition. In the ideal case, we examine the potential of our technique to modify the F0 patterns under a condition where the original F0 contours of test sentences are known. In the practical case, we show how the users intuitively modify the pitch by changing the initial F0 context labels obtained from the input text.

This paper describes an approach to HMM-based expressive speech synthesis which does not require any supervised labeling process for emphasis context. We use appealing-style speech whose sentences were taken from real domains. To reduce the cost for labeling speech data with an emphasis context for the model training, we propose an unsupervised labeling technique of the emphasis context based on the difference between original and generated F0 patterns of training sentences. Although the criterion for the emphasis labeling is quite simple, subjective evaluation results reveal that the unsupervised labeling is comparable to the labeling conducted carefully by a human in terms of speech naturalness and emphasis reproducibility. Index Terms: HMM-based speech synthesis, expressive speech, emphasis expression, unsupervised labeling, F0 generation

This paper proposes a speech segment selection method based on machine learning for concatenative speech synthesis systems. The proposed method has two novel features. One is its use of support vector machine (SVM) to estimate the subjective correctness of pitch accent with respect to each accent phrase of possible candidate speech segments. The other is its use of a determination function to identify the best segment based on SVM output. The determination function involves two assessments; one counts the number of each sign of SVM output and the other compares the distance values. The sign of SVM output is generally used to classify target objects, but the distance SVM output also represents important information. An experiment that assesses SVM performance for Japanese accent validity shows that its accuracy is 81%. To confirm the effectiveness of the proposed segment selection method, preference tests are conducted. The test indicates that the proposed method can yield Japanese synthesized speech with more natural intonation than the conventional method that uses only target cost and concatenation cost.

This paper describes a new recording script generation method that can create speech databases for corpus-based TTS (text to speech) systems. This method is efficient due to its two features: (1) it has a 2-stage algorithm to generate the recording script with consideration of the balance of triphone, syllable and morpheme elements; (2) it can control types of phonetic elements included in the recording script via the weight coefficients of the phonetic elements. An evaluation shows that the 2-stage algorithm is effective in raising the coverage of phonetic elements and that this method yields a recording script containing various phonetic elements. A preference test shows that changing the selection criteria influences the quality of the synthesized speech. The same test also shows that it is better to take account of morpheme-based elements than syllable-based elements in generating a task-specific recording script.

Modern Japanese texts often include Western sourced words written in the Roman alphabet. Even Japanese sourced words are sometimes transliterated into the Roman alphabet. As most of them are very new and idiosyncratic proper nouns, it is impractical to assume all those alphabetic words can be registered in the word dictionary of a text-to-speech system; their pronunciation must be derived automatically. As long vowel expressions are the same with short vowel or diphthong expressions in Japanese words transliterated into the alphabet, long vowel detection is necessary for generating highly accurate pronunciation. This paper proposes a method to detect long vowel positions using the Support Vector Machine. The cost of making the learning model for the detection is very low because the training data can be generated automatically from the text-to-speech word dictionary. 93.2% word accuracy was achieved which is almost the maximum accuracy possible from just spelling information.

This paper proposes a new text-to-speech (TTS) system that utilizes large numbers of speech segments to produce very natural and intelligible synthetic speech. There are two innovations; new multiform synthesis units and a new speech modification algorithm based on a vocoder that offers harmonics reconstruction. The multiform units make it possible to reduce acoustic discontinuities at concatenation points and unnatural sound by preparing synthesis units with various lengths and various F/sub 0/ contours. The new speech modification algorithm, on the other hand, improves the quality of prosody modified speech. This algorithm is extremely effective in synthesizing speech whose prosodic parameters are quite different from those of synthesis units. Listening tests confirm that the new synthesis units yield speech with high intelligibility and naturalness, and that the new speech modification algorithm is superior to all other conventional vocoders and waveform domain algorithms including TD-PSOLA, especially when modifying the F/sub 0/ frequency upward.

This paper describes a new framework for designing speaker recognition systems based on the discriminative feature extraction (DFE) method. We apply a mel-cepstral estimation technique to the feature extractor in a Gaussian mixture model (GMM)-based text-independent speaker identification system. The mel-cepstral estimation technique uses the second-order all-pass warping function for frequency transformation. We jointly optimize the frequency warping parameters of the feature extractor and the GMM parameters of the classifier based on a minimum classification error (MCE) criterion. Experimental results show that the frequency warped scale after optimization is different from traditional linear/mel scales; moreover, the proposed system outperforms conventional systems trained with the generalized probabilistic descent (GPD) method in which only the classifier is optimized.

In this paper, we first introduce the use of Gaussian mixture models (GMM) for Chinese tone classification in continuous speech. Then, we explain how to integrate it with the HMM-based speech recognition system. Finally, we provide the tone classification accuracy of this probabilistic method which is tested with Chinese continuous speech database of national “863” project.

To facilitate multimedia content production for Internet distribution, this paper proposes a new framework for generating multimedia content. Advantages of the framework are minimized amount of transmitted data and easy-to-update multimedia content. Key points are (1) to produce speech messages by a text-to-speech (TTS) system and (2) to produce moving pictures by concatenating templates from a set of moving pictures. To support the proposed framework, a speech design tool and a movie assign tool are introduced. Three kinds of contents have been produced using the proposed framework and informally evaluated. Important results are: (1) the transmitted data of WebMessenger is 120 times smaller than the full movie version of the content, (2) quality of speech synthesized by TTS is not sufficient and a speech design tool is necessary, (3) WebMessenger is 50% cheaper than human speech recording, (4) speech synthesis by combining manual and automatic methods is effective.

The paper proposes a new algorithm to convert the speech of one speaker so that it sounds like that of another. To modify formant frequencies and spectral tilt, the algorithm provides piecewise linear conversion rules which make it possible to control spectrum parameters in detail. Because the rules are statistically generated using training data, reliability of the conversion rules are guaranteed, and the rules are automatically generated for any pair of speakers. According to the conversion rules, formant frequencies and spectral tilt are modified by a formant modification algorithm based on the minimization of spectral distortion. Evaluation by listening tests shows that the proposed algorithm makes it possible to convert speaker individuality while maintaining high quality.>