FAQs



Q. For sub-task 1, baseline recipe is Kaldi based. Can participants use other model, e.g., end-to-end?

A. Yes.


Q. Can data from one subtask be used for the other subtask?

A. Yes.


Q. Can acoustic data other than the data released for this challenge be used for any of the subtasks?

A. Yes (not violating Challenge Rules), as long as the details of the technique or resources are made publicly available.


Q. How many trial submissions are allowed for the blind test set?

A. Five. If more than five are submitted, sixth and above submissions will not be accepted.


Q. Are participants allowed to use any data-augmentation technique on the released data, for example, speech rate modification, noise addition etc.?

A. Yes (not violating Challenge Rules), as long as the details of the technique or resources are made publicly available.


Q. Are participants allowed to use pre-trained acoustic or language models or any other models in system building?

A. Yes. Please refer to Challenge Rules.


Q. Can a participant submit more than one model's hypothesis? Will that count for the max five submission? If a participant wants to try two different models, will five submissions from each of them be allowed?

A. No, overall five submissions are allowed irrespective of how many models a participant tries on his/her end.


Q. If a group of individuals is participating in the challenge as a team do they mention everyone's details in the name and email address field (during registration) or just one of us would suffice?

A. Just one will do. We suggest to mention these details of the key contact person of the team.


Q. For code-mixing challenge, for English , can we use external English ASR data, example - Librispeech as additional data?

A. Yes (not violating Challenge Rules, as long as the data is (made) publicly available.


Q. If someone wants to use transliteration for some model, can he/she use an open source transliteration tool, which is already published elsewhere and available on GitHub publicly or that might be trained on external data?

A. Yes (not violating Challenge Rules), as long as the details of the technique or resources are (made) publicly available.


Q. Can a participant build his/her own lexicon?

A. Yes (not violating Challenge Rules), as long as the details of the technique or resources are (made) publicly available.


Q. Will the challenge participants of the sub-task1 be informed from which language each of the blind test set audios are taken?

A. No. Sub-task1 is on multilingual ASR. Hence, language information for the blind test set won’t be provided.


Q. For subtask-1, can participants use multiple ASR models , i.e. one trained ASR model per language for all the six languages given? Do these models need to be automatically selected (i.e. maybe using another deep neural network model) as per the language?

A. Yes. as long as the details of the technique are made available.


Q. For subtask-1 , is pseudo-labelling allowed once the test data-sets become available?

A. Yes, as long as publicly domain models are used (not violating challenge rules), not using services like Google ASR or other proprietary systems. Need to provide details of external resources and details of the technique used.


Q. For subtask-1, can participants use multiple ASR models , i.e. one trained ASR model per language for all the six languages given? Do these models need to be automatically selected (i.e. maybe using another deep neural network model) as per the language?

A. Yes. as long as the details of the technique are made available.


Q. For subtask-1 , is pseudo-labelling allowed once the test data-sets become available?

A. Yes, as long as publicly domain models are used (not violating challenge rules), not using services like Google ASR or other proprietary systems. Need to provide details of external resources and details of the technique used.


Q. For subtask-1, are participants allowed to use individual language models for each language?

A. Yes. as long as the details of the technique are made available.


Q. For subtask-2, can participants use individual ASR models for the two code-switched tasks?

A. Yes, this is allowed.