Zdeněk Hanzlíček

Czech HMM-based speech synthesis

In this paper, first experiments on statistical parametric HMM-based speech synthesis for the Czech language are described. In this synthesis method, trajectories of speech parameters are generated from the trained hidden Markov models. A final speech waveform is synthesized from those speech parameters. In our experiments, spectral properties were represented by mel cepstrum coefficients. For the waveform synthesis, the corresponding MLSA filter excited by pulses or noise was utilized. Beside that basic setup, a high-quality analysis/synthesis system STRAIGHT was employed for more sophisticated speech representation. For a more robust model parameter estimation, HMMs are clustered by using decision tree-based context clustering algorithm. For this purpose, phonetic and prosodic contextual factors proposed for the Czech language are taken into account. The created clustering trees are also employed for synthesis of speech units unseen within the training stage. The evaluation by subjective listening tests showed that speech produced by the combination of HMM-based TTS system and STRAIGHT is of comparable quality as speech synthesised by the unit selection TTS system trained from the same speech data.

Robust Methodology for TTS Enhancement Evaluation

The paper points to problematic and usually neglected aspects of using listening tests for TTS evaluation. It shows that simple random selection of phrases to be listened to may not cover those cases which are relevant to the evaluated TTS system. Also, it shows that a reliable phrase set cannot be chosen without a deeper knowledge of the distribution of differences in synthetic speech, which are obtained by comparing the output generated by an evaluated TTS system to what stands as a baseline system. Having such knowledge, the method able to evaluate the reliability of listening tests, as related to the estimation of possible invalidity of listening results-derived conclusion, is proposed here and demonstrated on real examples.

Experiments on Reducing Footprint of Unit Selection TTS System

The quality of speech produced by modern TTS systems utilizing the unit selection approach is very high. However, the system demands are enormous. The storage requirements are directly proportional to the size of speech unit inventory from which the units are selected during the synthesis process. This paper presents the analysis and reduction experiments performed on two large speech corpora employed by a unit selection TTS system for the Czech language. A procedure for exclusion of utterances from the default speech corpus based on statistics of the usage of particular speech units was proposed. The exclusion of whole utterances from the corpus was preferred over the exclusion of individual speech units in order to preserve the fundamental feature of the unit selection method – selection of possibly longest sequences of speech units. Experiments were performed for several reduction levels. Resulting synthetic speech was evaluated by a proposed statistics based on the concatenation points density. Moreover, the speech quality was evaluated in listening tests. All reduced versions of TTS system were evaluated as similar or slightly worse than the baseline system.

Voice Conservation: Towards Creating a Speech-Aid System for Total Laryngectomees

This paper describes the initial experiments on voice conservation of patients with laryngeal cancer in an advanced stage. The final aim is to create a speechaid device which is able to “speak” with their former voices. Our initial work is focused on applicability of speech data from patients with an impaired vocal tract for the purposes of speech synthesis. Preliminary results indicate that appropriately selected synthesis method can successfully learn a new voice, even from speech data which is of a lower quality.

Current State of Text-to-Speech System ARTIC: A Decade of Research on the Field of Speech Technologies

This paper provides a survey of the current state of ARTIC – the modern Czech concatenative corpus-based text-to-speech system. Through more than a decade of research & development in the field of speech technologies and applications, the system was enriched with new languages (and, as a consequence, language-dependent NLP methods), and its speech generation capabilities were significantly improved when new progressive speech generation modules (SPS, DNN, HSS) were (and are still being to) designed and incorporated into it. Also, ARTIC has to deal with various requirements on data used to generate speech from, ranging in size, quality and domain of the output speech, while there always was the requirement to achieve the highest quality in terms of both naturalness and intelligibility. Thus, the paper summarizes some of the most significant achievements and demanding tasks which had to be tackled by the system, illustrating the universality and flexibility of this Czech TTS system.

Czech Expressive Speech Synthesis in Limited Domain Comparison of Unit Selection and HMM-Based Approaches ∗

This paper deals with expressive speech synthesis in a limited domain restricted to conversations between humans and a computer on a given topic. Two different methods (unit selection and HMM-based speech synthesis) were employed to produce expressive synthetic speech, both with the same description of expressivity by so-called communicative functions. Such a discrete division is related to our limited domain and it is not intended to be a general solution for expressivity description. Resulting synthetic speech was presented to listeners within a web-based listening test to evaluate whether the expressivity is perceived as expected. The comparison of both methods is also shown.

Correction of Prosodic Phrases in Large Speech Corpora

Nowadays, in many speech processing tasks, such as speech recognition and synthesis, really large speech corpora are utilized. These speech corpora usually contain several hours of speech or even more. To achieve possibly best results, an appropriate annotation of the recorded utterances is often necessary. This paper is focused on problems related to the prosodic annotation of the Czech speech corpora. In the Czech language, the utterances are supposed to be split by pauses into so-called prosodic clauses containing one or more prosodic phrases. The types of particular phrases are linked to their last prosodic words corresponding to various functionally involved prosodemes. The clause/phrase structure is substantially determined by the sentence composition. However, in real speech data, different prosodeme type or even phrase/clause borders can be present. This paper deals with 2 basic problems: the correction of the improper prosodeme/phrase type and the detection of new phrase borders. For both tasks, we proposed new procedures utilizing hidden Markov models. Experiments were performed on 4 large speech corpora recorded by professional speakers for the purpose of speech synthesis. These experiments were limited to the declarative sentences. The results were successfully verified by listening tests.

Initial experiments on automatic correction of prosodic annotation of large speech corpora

Most modern speech synthesis systems utilize large speech corpora to learn new voices. These speech corpora usually contain several hours of speech spoken by talented speakers who are able to record such an amount of speech data in a sufficient quality. An appropriate phonetic and prosodic annotation of the recorded utterances is necessary for a high quality of synthesized speech. For many languages, the pitch shape within the last prosodic word of a phrase is characteristic for particular types of sentences and phrase structure of compound/complex sentences. However in the real data, this formal convention can be breached and a different pitch shape than expected can be present. This can be a source of prosody inconsistency in synthesized speech. This article presents some experiments on automatic detection of prosodic mismatch in recorded utterances. A simple classifier based on GMM was proposed for this task. Experiments were performed on 5 large speech corpora. The classification results were successfully verified by listening tests.

Czech HMM-based speech synthesis: experiments with model adaptation

This paper describes some experiments on model adaptation for statistical parametric speech synthesis for the Czech language. For building an experimental TTS system, HTS toolkit was utilised. Speech was represented by using high-quality analysis/synthesis system STRAIGHT. For definition of speech unit context, a new reduced set of contextual factors was proposed. During model clustering, some missing contextual factors, that were not included in this set, can be simulated by using combined context-related clustering questions. The model transformation was performed by a combination of CMLLR and MAP adaptation. Speech data from 3 male and 3 female speakers was used in our experiments. In the performed listening test, speech generated from regularly trained and adapted models was compared. Both voices were evaluated as identical and of a similar quality.

Optimal Number of States in HMM-Based Speech Synthesis

This paper deals with using models with a variable number of states in the HMM-based speech synthesis system. The paper also includes some implementation details on how to use these models in systems based on the HTS toolkit, which cannot handle the models with an unequal number of states directly. A bypass to enable this functionality is proposed here. A data-based method for the determination of the optimal number of states for particular models is proposed here and experimentally tested on 4 large speech corpora. The preference listening test, focused on local differences, proved the preference of the proposed system to the traditional system with 5-state models, while the size of the proposed system (the total number of states) is lower.