Tamil Speech Recognition Using XLSR Wav2Vec2.0 & CTC Algorithm

Automatic Speech Recognition is a promising research topic with lots of real-world applications like virtual assistants, aids for physically challenged etc. Tamil language speech recognition could be potentially challenging due to the fact that there are many possible dialects, slangs and accents. This paper proposes an ASR system based on cross-lingual transfer learning in combination with CTC algorithm. The pretrained model from Facebook AI viz. XLSR Wav2Vec2.0 is used. The dataset used in this work is Common Voice Tamil, which is a crowd-sourced dataset provided by Mozilla. Our system achieves a Word Error Rate of 0.58 and Character Error Rate of 0.11.