Acoustic modeling using an extended phone set considering cross-lingual pronunciation variations.