The CSTR entry to the Blizzard Challenge 2021

Dan Wells, Pilar Oplustil-Gallegos, Simon King
The Centre for Speech Technology Research
University of Edinburgh

This is the companion webpage for our entry to the Blizzard Challenge 2021 spoke task for synthesizing Spanish utterances including some English words. For a detailed description of the system we submitted, read our paper here.

Audio samples

The table below presents a selection of the audio samples submitted to the main Blizzard Challenge evaluation and used in our supplementary listening test. In our experiments with different nativization strategies, each sample was presented alongside its text transcript, with English words highlighted in bold.

We asked participants to rate each sample on a 5-point Likert scale (from "disagree" to "agree") with regard to the following statements:

1. This phrase is very natural.
2. Considering the marked words, this speaker shows mastery of English.
3. I would pronounce the English words the same way.

System key

• Blizzard – System submitted for Blizzard evaluation: phonological feature inputs, pronunciations and word-level language embeddings determined by automatic language ID
• Mix – Pronunciations and word-level language embeddings determined by ground-truth language labels
• Nat – Pronunciations determined by ground-truth language labels, all language embeddings set to Spanish (nativized)
• Esp – All pronunciations produced by Spanish g2p rules only (no nativization of English words) and all language embeddings set to Spanish
• F – Phonological feature inputs
• P – Phoneme inputs