SPRING LAB MARATHI-STREAMING

About Model

Automatic Speech Recognition (ASR) model forMarathi speech recognition, developed using the Icefall toolkit with the Zipformer architecture. The model is trained on a dataset consisting of approximately 400 hours of labelled speech. It is trained on 16 kHz audio, including naturally occurring code-mixed speech, enabling robust recognition of bilingual Indian speech patterns. The system is based on a 65M-parameter Zipformer-Medium encoder, paired with an RNN-T prediction network and joiner, forming a low-latency streaming ASR model with 16 encoder layers and a 512-dimensional representation.