Anandaswarup
Vadapalli
← All projects

SPEECH & LANGUAGE / 2013 · ACL

English words in Telugu speech

Comparing word-to-phone and phone-to-phone mapping for English words in a Telugu text-to-speech system.

01 / MOTIVATION

The question

Is word-to-phone mapping better than phone-to-phone mapping for handling English words?

02 / METHOD

The approach

The ACL 2013 study compares English pronunciation strategies using only the native Telugu phone set. An initial comparison uses manually prepared mappings for 10 bilingual utterances and listening evaluations by five native Telugu speakers.

The work then learns word-to-phone mappings using automatic epsilon scattering for alignment and classification and regression trees. The learned mappings are evaluated on held-out and different-domain data and incorporated into a Telugu TTS system.

03 / EVIDENCE

Findings & scope

In the initial manual-mapping comparison, word-to-phone mapping received a mean opinion score of 3.48 versus 2.66 for phone-to-phone mapping. It was preferred in 32 of 50 ABX judgments, compared with 4 for phone-to-phone mapping and 14 with no preference.

The initial listener comparison is small and specific to Telugu. The paper separately evaluates learned mappings, so the manual-mapping result should not be interpreted as the performance of the automated system.

READ THE WORK

Related publications

2013

Is word-to-phone mapping better than phone-phone mapping for handling English words?

Naresh Kumar Elluru, Anandaswarup Vadapalli, Raghavendra Elluru, Hema A. Murthy, Kishore Prahallad

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, ACL 2013, 4-9 August 2013, Sofia, Bulgaria, Volume 2: Short Papers

NEXT PROJECTLearning patterns in hospital readmission→