An evaluation set for speech recognition experiments used in [1]. This set contains 6 interview sessions (close talk microphone) of Singaporean students reading passages about Singapore streets. Each interview sessions consists of 2 recordings corresponding to interviewer and interviewee.

• Compound street names are combined using underscore symbol, i.e. 'boon lay' -> 'boon_lay'.
• To simulate the scenario where these street names are rare words, ensure that they are absent or appear a few times (1-3) in the train set.
• The main application of the evaluation set is to correctly recognize these street names while preserving the WER.
• List of named entities are given in file named_entities.txt.
• List of other Singapore street names can be found in: https://geographic.org/streetview/singapore/
• The pronunciation lexicon can be obtained from G2P models such as http://www.speech.cs.cmu.edu/tools/lextool.html. E.g. for word "boon_lay", first generate pronunciation lexicon of "boon" and "lay" separately, and then combine all pronunciation variations.
• The readers are Singaporeans (accent is different from English speakers in other countries), and thus train set containing Singapore English is recommended, e.g. https://www2.imda.gov.sg/NationalSpeechCorpus. (Note: in the paper we used strong Singapore English ASR trained on our own dataset).

[1] Khassanov, Y., Zeng, Z., Pham, V.T., Xu, H., Chng, E.S. (2019) Enriching Rare Word Representations in Neural Language Models by Embedding Matrix Augmentation. Proc. Interspeech 2019, 3505-3509, DOI: 10.21437/Interspeech.2019-1858..