We will invite participants of this Special Session to build their own ASR systems for Indian languages, in the broad sense of the term. You may choose which Indian languages to build systems for, and what data sets to use.

When building these systems, participants could consider:

• mixing together audio and text data from multiple languages, e.g. building a Hindi + Urdu system, or building one end-to-end system covering multiple languages
• folding in text corpora from more sources than the training data (e.g. from a web crawl run by the participants)

If the technical approach taken by the participants requires pronunciation lexicons, they would need to create these themselves.

Individual participants will be responsible for evaluating their machine transcripts on their test sets. They should write a paper explaining their set-up, research ideas, and key innovations.

Since we also want to stimulate research on how ASR may be applied for Indian languages, Google will be providing a limited number of Cloud Speech API credits to participants, who can then build voice-enabled apps using ASR for Indian languages. (Participants will not be required to use Google's Cloud Speech API, though.)

List of Resources Needed

• ASR training data for some Indian languages (See examples in Data.)
• Cloud Speech API credits for show-and-tell (a limited number to be provided by Google. Contact: interspeech2018-ss1-committee@google.com)