This notebook contains:

• Notebooks finetuning, inferencing and generating N-gram.
• Demo Whisper and Wav2vec

For using Beam search with LM, install Whisper from my Github

    pip install git+https://github.com/HKAB/whisper.git

For training & inference

  python finetune.py  --model_name base \
                      --dataset vin100h

  python test.py      --checkpoint_path path/to/ckpt \
                      --dataset vin100h \
                      --model_name base

For generating language model with KenLM, use notebook in notebooks folder.

We share the checkpoint (base, batch_size 1, gradient accumulation steps 10, epoch 14, lr 0.0001).

• We finetune Whisper on 100h speech dataset.
• We implement BeamSearchWithLM, using KenLM and showed positive result.