Logo Lanfrica

anzeyimana/KinyaEmformerRNNT

Domaine:

natural language processing

Type de record:

softwaremodel
Créateur:
anz
Hôte:
Kinyarwanda streaming ASR based on Emformer RNN-T model architecture # KinyaEmformerRNNT This codebase implements a Emformer model training and inference scripts for lightweight Kinyarwanda streaming speech recognition. The model can be deployed on a smartphone for real-time speech recognition using the phone's micrcophone. ## Getting started TODO: We will be releasing the pre-trained model for Kinyarwanda soon for inference on Android. ### Inference Generating TorchScript model to be deployed on Android ```` cd Inference/ python generate_ts.py python save_model_for_mobile.py ```` ### Training Training script using Mozilla Common Voice Kinyarwanda dataset: ```` cd Training/emformer_rnnt/ python train_cv.py ````