Discovering Canonical Indian English Accents: A Crowdsourcing-based Approach

Discovering Canonical Indian English Accents: A Crowdsourcing-based Approach Sunayana Sitaram∗, Varun Manjunath†, Varun Bharadwaj†, Monojit Choudhury∗, Kalika Bali∗, Michael Tjalve+ ∗Microsoft Research, Bangalore, India †PES University, Bangalore, India susitara, monojitc, kalikab@microsoft.com varunmanjunath2012, varunbharadwaj1995@gmail.com +Microsoft Corporation, Redmond, USA University of Washington, Seattle, USA mitjalve@microsoft.com Abstract Automatic Speech Recognition (ASR) systems typically degrade in performance when recognizing an accent different from the accents in the training data. One way to overcome this problem without training new models for every accent is adaptation. India has over a hundred major languages, which leads to many variants in Indian English accents. Making an ASR system work well for Indian English would involve collecting data for all representative accents in Indian English and then adapting Acoustic Models for each of those accents. However, given the number of languages that exist in India and the lack of a prior work in literature about how many Indian English accents exist, it is difficult to come up with a set of canonical accents that could sufficiently capture the variations observed in Indian English. In addition, there is a lack of labeled corpora of accents in Indian English. We approach the problem of determining a set of canonical Indian English accents by taking a crowdsourcing based approach. We conduct a mobile app based user study in which we play audio samples collected from all over India and ask users to identify the geographical origin of the speaker. We measure the consensus among users to come up with a set of candidate accents in Indian English and identify which accents are best recognized and which ones are confusable. We extend our preliminary user study to a web app-based study that can potentially generate more labeled data for Indian English accents. We describe results and challenges encountered in a pilot study conducted using the web-app and future work to scale up the study. Keywords: non-native accents, crowdsourcing, Indian languages, speech processing 1. Introduction Automatic Speech Recognition (ASR) systems have reached performance on par with humans on some tasks and languages. However, the performance of ASR systems is significantly worse in the presence of accents that are different from the accents in the data used for training the ASR system. Since commercial ASR systems require hun- dreds or thousands of hours to train, it is not always fea- sible to train a separate system for every accent the ASR system will have to recognize. One solution to this is to adapt Acoustic Models that are trained on one accent to a target accent by using a small amount of speech data in the target accent (Huang et al., 2014). India is an extremely multilingual country. According to the 2001 Census of India (Banthia, 2001) there are 122 ma- jor and 1599 other languages in India; 23 languages have been granted the status of official languages. The existing literature on Indian English (IE) accents either focuses on its difference with British or American English, or on very specific phonetic features that mark IE (Sahgal …

Content truncated to 500 words