Hypemasality detection in cleft palate speech based on natural computation

The speakers with cleft palate, due to the defective velopharyngeal mechanism, allow the passage of air through the nasal cavity, which introduces inappropriate nasal resonance during speech production and results in hypernasal speech. The existence of hypernasality severely reduces the intelligibility of the speech. The treatment of cleft palate hypernasal speech requires the follow up operation to close fracture and restore the normal voice. Speech evaluation is essential to assess the hypernasality grades. In this work, an automatic hypernasality grades detection method is proposed in cleft palate speech. After a low quefrency liftering at 90 quefrencies cutoff in cepstrum domain, a homomorphic spectrum is calculated as the extraction feature. Then a BP neural network classifier based on natural computation is applied to detect four grades of hypernasality: normal, mild, moderate and severe. The experiment results show that the classification accuracy for four grades of hypernasality is above 80%.