The details of the implementation of a Marathi Automatic Speech Recognition (ASR) system as well as the associated speech database is given in this paper. The speech database consists of more than 15000 speech files. Over 1500 speakers read 10 sentences each into their mobile phone. Speech was recorded over telephone channel. The text corpus consists of 3400 sentences consisting of over 10000 unique words. The ASR system was implemented using Kaldi toolkit. Acoustic models of various characteristics were implemented and 3-fold validation tests were conducted. The word error rate of recognising test data is 24%. Experiments were conducted to study the effect, on the performance of the system, of (a) manual annotation of non-speech events such as cough, babble etc., and (b) discarding of those training speech files which could not be recognised well. The results of these experiments and the lessons learnt are presented.