Building and evaluating synthetic voices

01 / MOTIVATION

The question

How can shared speech data and listener evaluations support meaningful comparisons between synthesis systems across languages and speaking styles?

02 / METHOD

The approach

The 2013, 2014, and 2015 challenge overview papers describe Indian-language tasks and listening evaluations. These papers document shared evaluation efforts, distinct from the IIIT Hyderabad system submission papers.

The 2015 IIIT Hyderabad submission uses unit-selection synthesis for monolingual and multilingual tasks in six Indian languages, with sentence-level Viterbi search and correlation-based overlap addition.

The 2017 submission examines neural acoustic models for expressive children’s audiobook synthesis. A BLSTM system is selected for submission following held-out objective comparisons with a hybrid architecture.

03 / EVIDENCE

Findings & scope

Together, the papers document both the design of shared evaluations and the construction of synthesis systems. The 2017 paper reports improvements over DNN and HMM baselines in most evaluations.

Tasks, languages, and systems differ across challenge years. The collection should be read as a research progression, not as a single comparable performance series.