Sequence-to-Sequence Acoustic-to-Phonetic Conversion Using Spectrograms and Deep Learning

Distinctive phonetic features (DPFs) abstractedly describe the place, manner of articulation, and voicing language phonemes. While DPFs are powerful speech signals that capture unique articulatory characteristics each phoneme, task DPF extraction is challenged by need for efficient computational model. Unlike ordinary acoustic can be directly determined form waveform using closed-form expressions, elements extracted from machine learning (ML) techniques. Therefore, objective developing an acoustic-to-phonetic converter high accuracy low complexity, it important to select input simple, yet carry adequate information. This paper examines effectiveness spectrogram as feature with modeled two deep techniques: belief network (DBN) convolutional recurrent neural (CRNN). The proposed method applied on Modern Standard Arabic (MSA). Multi-label modeling considered in converter. techniques were evaluated proper evaluation measures accommodate imbalanced nature elements. results showed CRNN more accurate extracting than DBN.