Blizzard Challenge 2017
Read these first
This year, there are two distinct parts to the Blizzard Challenge. Teams may enter either one, or both. The first part of the challenge follows the standard approach of previous years, and comprises the single hub task (2017-EH1) which requires teams to build an end-to-end text-to-speech system. The second part of the challenge is novel and is designed to be accessible to the wider machine learning community; it comprises two spoke tasks (2017-ES1 and 2017-ES2)
- First, please read the calls for participation:
- Before participating, please read and agree to the rules for whichever part(s) of the challenge you are interested in
- You should only register for the challenge if you actually plan to submit an entry to the challenge
New: the Blizzard Machine Learning Challenge
Speech synthesis as a machine learning problem ---exploring new types of acoustic models
In the HMM era, by taking a unified view of both Automatic Speech Recognition (ASR) and Text-to-Speech (TTS), it was possible to develop various types of new ASR and TTS techniques, e.g., cross-lingual speaker adaptation, adaptive training for TTS, use of prosody in ASR, etc. We expect that by once again taking a unified view in the current DNN era, it will be possible to develop new types of acoustic modeling techniques that are useful for both ASR and TTS.
The series of Blizzard Challenges has helped us measure progress in TTS. But, to get competitive performance, a lot time has to be spent on skilled tasks such as updating the lexicon, removing inappropriate audio files, segmenting and aligning audio files, detecting alignment errors, etc. This may make the Blizzard Challenge unattractive to Machine Learning (ML) researchers from other fields.
We therefore propose a spin-off challenge that does not involve these speech-specific tasks, and allows participants to concentrate on the acoustic modeling task, framed as a straightforward ML problem, with a fixed data set.
The data that the organizers will provide is in the form of corresponding sequences of linguistic features, speech features and speech waveforms. Participants must train a model to predict speech features from linguistic features (or, to directly predict speech waveforms from linguistic features, as done in WaveNet), and then use that model to make predictions for a test set of previously-unseen linguistic features.
Evaluation will be done by the organisers, using a listening test, as in the main challenge.
Registration
Register by emailing blizzard@festvox.org. We need to know your team name, the name of the main contact person, your affiliation, and contact details including email address, postal address and phone number. Please specify which task(s) you plan to submit entries for.
Data download
The speech + text data comes from professional audiobooks produced by Usborne Publishing.
- 2017-EH1
- About 6.5 hours of British English speech data from a single female talker, which comprises 5 hours of speech already released for the 2016 challenge plus the audio from 6 additional books that were used for test material in 2016.
- Processed versions, such as alignments, are shared via the Blizzard Challenge 2016-7 Git Repository
- 2017-ES1 and 2017-ES2
- About 5 hours of British English speech data (waveforms) from a single female talker, as used in the 2016 challenge, along with linguistic features and speech features.
Download links can be found via http://www.cstr.ed.ac.uk/projects/blizzard/2017/usborne_blizzard2017
Development tools
- available via : http://www.cstr.ed.ac.uk/projects/blizzard/, which include submitted synthetic speech and listener scores for some previous Blizzard Challenges, which may be helpful during development
Questionnaire
- Download the questionnaire, complete it, and return it at the same time as your synthetic speech: http://data.cstr.ed.ac.uk/blizzard2017/system_questionnaire.txt
Mailing list
There is a mailing list for discussion and announcements for the challenge:
blizzard-discuss@festvox.org
Participants must join the list by sending a message to majordomo@festvox.org with the following line in the body of the message
subscribe blizzard-discuss
Once you are a member you will be able to mail messages to blizzard-discuss@festvox.org
Timeline
The timeline shown on this web page is the official one and supercedes those shown in announcements - it is subject to change, but we will try to follow it as closely as possible. Note that we will not consider any requests from participants to change the synthetic speech submission date or the paper submission date!
Dec 8 2016 - 2017-EH1 database released Jan ? 2016 - 2017-ES1 and 2017-ES2 database released Mar 2017 - test sentences released to participants Apr 2017 - participants submit their output, plus questionnaire (by midnight UTC) Apr 2017 - evaluation systems go live Jun 2017 - end of evaluation period Jun 15 2017 - release of results Jun 29 2017 - deadline to submit workshop papers (for either workshop) Jul 31 2017 - notification of paper acceptance for 2017-EH1 Aug 20-24 2017 - Interspeech 2017, Stockholm, Sweden Aug 25 2017 - Blizzard Challenge workshop, Stockholm - for task 2017-EH1 (date and location provisional) Aug 28- 2017 - EUSIPCO 2017, Kos, Greece Aug 31 2017 - notification of paper acceptance for 2017-ES1 and 2017-ES2 Dec 16-20 2017 - ASRU 2017 will include the workshop for tasks 2017-ES1 and 2017-ES2
Workshop
Information on the two workshops can be found here:
Any questions?
- Please contact blizzard@festvox.org if you have any questions
Previous challenges
- Blizzard Challenge 2016
- Blizzard Challenge 2015
- Blizzard Challenge 2014
- Blizzard Challenge 2013
- Blizzard Challenge 2012
- Blizzard Challenge 2011
- Blizzard Challenge 2010
- Blizzard Challenge 2009
- Blizzard Challenge 2008
- Blizzard Challenge 2007
- Blizzard Challenge 2006 papers and results: http://www.festvox.org/blizzard/blizzard2006.html
- Blizzard Challenge 2005 papers and results: http://www.festvox.org/blizzard/blizzard2005.html
