hyperparams.yaml

# Generated 2023-06-24 from:
# /netscratch/sagar/thesis/speechbrain/recipes/RescueSpeech/ASR/transformer/hparams/train_hf_whisper.yaml
# yamllint disable
# ################################
# Model: Whisper (Encoder-Decoder) + NLL
# Augmentation: TimeDomainSpecAugment
# Authors: Sangeet Sagar 2022
# ################################


# URL for the biggest Fairseq english whisper model.
whisper_hub: openai/whisper-large-v2
language: german

# Normalize the english inputs with
# the same normalization done in the paper
normalized_transcripts: true
test_only: false # Set it to True if you only want to  do the evaluation

auto_mix_prec: false
sample_rate: 16000

# These values are only used for the searchers.
# They needs to be hardcoded and should not be changed with Whisper.
# They are used as part of the searching process.
# The bos token of the searcher will be timestamp_index
# and will be concatenated with the bos, language and task tokens.
timestamp_index: 50363
eos_index: 50257
bos_index: 50258

# Decoding parameters
min_decode_ratio: 0.0
max_decode_ratio: 1.0
test_beam_size: 8

# Model parameters
freeze_whisper: false
freeze_encoder_only: false
freeze_encoder: true

#
# Functions and classes
#
whisper: &id001 !new:speechbrain.lobes.models.huggingface_whisper.HuggingFaceWhisper
  source: openai/whisper-large-v2/
  freeze: false
  save_path: openai/whisper-large-v2/
  encoder_only: false
  freeze_encoder: true


modules:
  whisper: *id001
whisper_opt_class: !name:torch.optim.AdamW
  lr: 0.00003
  weight_decay: 0.01

decoder: !new:speechbrain.decoders.seq2seq.S2SWhisperGreedySearch
  model: *id001
  bos_index: 50363
  eos_index: 50257
  min_decode_ratio: 0.0
  max_decode_ratio: 1.0

pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
  loadables:
    whisper: !ref <whisper>