Law cost training of hmm-based speech recognition or synthesis system with low size Arabic labeled unsegmented noisy speech database

Abstract EN

The first step in the process of developing speech recognition and formant or statistical parametric speech synthesis applications is preparation of speech training database.

The recognition and synthesis quality depend crucially on speech database size and fidelity.

But, preparing large speech training data for these applications is very costly in effort and time since its recorded waveform files should be segmented and labeled according to the class of labels used in the system.

This paper discusses a way to overcome the problem of training data starvation and the high cost of preparing it by introducing a method for training Hidden Markov Model (HMM)-based system with low size Arabic labeled unsegmented noisy speech database.

This method can be used for developing HMM-based recognition or HMM-based synthesis system in which speech waveform is generated from HMMs themselves.

Subjective evaluation of proposed Arabic speech HMM-based synthesis system shows that also increasing the length of context-dependent models makes contextual factors more prominent and improves the formant structure of the generated phonemes but will have the disadvantage of poor generalization and perceivable clicks in the synthesized speech.