How mdhaffar/named-entity-recognition works
Project type and purpose
• This is a modular machine-learning training pipeline for spoken language understanding. It trains (a) an Automatic Speech Recognition (ASR) model, (b) a Text-to-ASR-Embeddings (T2AE) model, and (c) a Named Entity Recognition (NLU/SLU) model, then stitches them into an end-to-end system. It leverages PyTorch and SpeechBrain, driven entirely by Python scripts and YAML configs.
File structure
• Top-level scripts:
– train_ASR.py (uses wav2vec_ASR.yaml)
– train_NLU_from_TTS_emb.py (uses NER.yaml)
• Config files (YAML):
– wav2vec_ASR.yaml → ASR model settings
– NER.yaml → NLU model settings
• README.md for usage
There are no subdirectories, so each script is a standalone module in the pipeline.
Key insights from README
• Step 1: Train ASR
• Step 2: Extract last-layer ASR embeddings + train T2AE via SpeechBrain’s TTS recipe
• Step 3: Train NLU on extracted embeddings
• Step 4: Integrate ASR and SLU into an end-to-end flow