models/sb_phonemizer/hyperparams.yaml

# Generated 2022-07-09 from:
# /notebooks/speechbrain/recipes/LibriSpeech/G2P/hparams/hparams_g2p_rnn.yaml
# yamllint disable
# ################################
# Model: LSTM (encoder) + GRU (decoder) (tokenized)
# Authors:
# Loren Lugosch & Mirco Ravanelli 2020
# Artem Ploujnikov 2021
# ################################

# Seed needs to be set at top of yaml, before objects with parameters are made
seed: 1234
__set_seed: !apply:torch.manual_seed [1234]


# Tokenizers
char_tokenize: false
char_token_type: unigram  # ["unigram", "bpe", "char"]
char_token_output: 512
char_token_wordwise: true
phn_tokenize: false
phn_token_type: unigram  # ["unigram", "bpe", "char"]
phn_token_output: 512  # index(blank/eos/bos/unk) = 0
phn_token_wordwise: true
character_coverage: 1.0


phonemes_count: 43
graphemes_count: 31
phonemes_enable_space: true

# Training Parameters
lexicon_epochs: 50
lexicon_ctc_epochs: 10
lexicon_limit_to_stop: 50                    # No stopping by default, can override
lexicon_limit_warmup: 50                    # No stopping by default, can override
sentence_epochs: 13
sentence_ctc_epochs: 10
sentence_limit_to_stop: 3
sentence_limit_warmup: 3
homograph_epochs: 50
homograph_ctc_epochs: 10
homograph_limit_to_stop: 5
homograph_limit_warmup: 10
lexicon_batch_size: 1024
sentence_batch_size: 32
homograph_batch_size: 32
ctc_weight: 0.5
homograph_loss_weight: 2.0
lr: 0.002
save_for_pretrained: true

# Model parameters
output_neurons: &id004 !apply:speechbrain.utils.hparams.choice

  value: false
  choices:
    true: 513
    false: 43

enc_num_embeddings: &id005 !apply:speechbrain.utils.hparams.choice
  value: false
  choices:
    true: 513
    false: 31

enc_dropout: 0.5
enc_neurons: 512
enc_num_layers: 4
dec_dropout: 0.5
dec_neurons: 512
dec_att_neurons: 256
dec_num_layers: 4
embedding_dim: 512

# Determines whether to use BOS (beginning-of-sequence) or EOS (end-of-sequence) tokens
# Available modes:
# raw: no BOS/EOS tokens are added
# bos: a beginning-of-sequence token is added
# eos: an end-of-sequence token is added
grapheme_sequence_mode: bos
phoneme_sequence_mode: bos


# Special Token information
bos_index: 0
eos_index: 1
blank_index: 2
unk_index: 2
token_space_index: 512


# Language Model
lm_emb_dim: 256 # dimension of the embeddings
lm_rnn_size: 512 # dimension of hidden layers
lm_layers: 2 # number of hidden layers
lm_output_neurons: 43

# Beam Searcher
use_language_model: false
beam_search_min_decode_ratio: 0
beam_search_max_decode_ratio: 1.0
beam_search_beam_size: 16
beam_search_beam_size_valid: 16
beam_search_eos_threshold: 10.0
beam_search_using_max_attn_shift: false
beam_search_max_attn_shift: 10
beam_search_coverage_penalty: 5.0
beam_search_lm_weight: 0.5
beam_search_ctc_weight_decode: 0.4
beam_search_temperature: 1.25
beam_search_temperature_lm: 1.0

# Word embeddings
use_word_emb: true
word_emb_model: bert-base-uncased
word_emb_dim: 768
word_emb_enc_dim: 256
word_emb_norm_type: batch

graphemes: &id028
- A
- B
- C
- D
- E
- F
- G
- H
- I
- J
- K
- L
- M
- N
- O
- P
- Q
- R
- S
- T
- U
- V
- W
- X
- Y
- Z
- "'"
- ' '

phonemes: &id001


- AA
- AE
- AH
- AO
- AW
- AY
- B
- CH
- D
- DH
- EH
- ER
- EY
- F
- G
- HH
- IH
- IY
- JH
- K
- L
- M
- N
- NG
- OW
- OY
- P
- R
- S
- SH
- T
- TH
- UH
- UW
- V
- W
- Y
- Z
- ZH
- ' '

enc_input_dim: &id003 !apply:speechbrain.lobes.models.g2p.model.input_dim
  use_word_emb: true
  word_emb_enc_dim: 256
  embedding_dim: 512


phn_char_map: &id002 !apply:speechbrain.lobes.models.g2p.dataio.build_token_char_map


# Models
  tokens: *id001
char_phn_map: &id023 !apply:speechbrain.lobes.models.g2p.dataio.flip_map
  map_dict: *id002
enc: &id006 !new:speechbrain.nnet.RNN.LSTM
  input_shape: [null, null, *id003]
  bidirectional: true
  hidden_size: 512
  num_layers: 4
  dropout: 0.5

lin: &id010 !new:speechbrain.nnet.linear.Linear
  input_size: 512
  n_neurons: *id004
  bias: false

ctc_lin: &id013 !new:speechbrain.nnet.linear.Linear
  input_size: 1024
  n_neurons: *id004
encoder_emb: &id007 !new:speechbrain.nnet.embedding.Embedding
  num_embeddings: *id005
  embedding_dim: 512

emb: &id008 !new:speechbrain.nnet.embedding.Embedding
  num_embeddings: *id004
  embedding_dim: 512

dec: &id009 !new:speechbrain.nnet.RNN.AttentionalRNNDecoder
  enc_dim: 1024
  input_size: 512
  rnn_type: gru
  attn_type: content
  dropout: 0.5
  hidden_size: 512
  attn_dim: 256
  num_layers: 4

word_emb_enc: &id012 !new:speechbrain.lobes.models.g2p.model.WordEmbeddingEncoder

  word_emb_dim: 768
  word_emb_enc_dim: 256
  norm_type: batch

word_emb: !apply:speechbrain.lobes.models.g2p.dataio.lazy_init
  init: !name:speechbrain.wordemb.transformer.TransformerWordEmbeddings
    model: bert-base-uncased

log_softmax: &id011 !new:speechbrain.nnet.activations.Softmax
  apply_log: true

modules:
  model: &id014 !new:speechbrain.lobes.models.g2p.model.AttentionSeq2Seq
    enc: *id006
    encoder_emb: *id007
    emb: *id008
    dec: *id009
    lin: *id010
    out: *id011
    use_word_emb: true
    word_emb_enc: *id012
  enc: *id006
  encoder_emb: *id007
  emb: *id008
  dec: *id009
  lin: *id010
  ctc_lin: *id013
  out: *id011
  word_emb:
  word_emb_enc: *id012
model: *id014
lm_model: &id015 !new:speechbrain.lobes.models.RNNLM.RNNLM
  embedding_dim: 256
  rnn_layers: 2
  rnn_neurons: 512
  output_neurons: 43
  return_hidden: true

opt_class: !name:torch.optim.Adam
  lr: 0.002

beam_searcher: &id029 !new:speechbrain.decoders.S2SRNNBeamSearcher
  embedding: *id008
  decoder: *id009
  linear: *id010
  ctc_linear: *id013
  bos_index: 0
  eos_index: 1
  blank_index: 2
  min_decode_ratio: 0
  max_decode_ratio: 1.0
  beam_size: 16
  eos_threshold: 10.0
  using_max_attn_shift: false
  max_attn_shift: 10
  coverage_penalty: 5.0
  ctc_weight: 0.4

beam_searcher_valid: !new:speechbrain.decoders.S2SRNNBeamSearcher
  embedding: *id008
  decoder: *id009
  linear: *id010
  ctc_linear: *id013
  bos_index: 0
  eos_index: 1
  blank_index: 2
  min_decode_ratio: 0
  max_decode_ratio: 1.0
  beam_size: 16
  eos_threshold: 10.0
  using_max_attn_shift: false
  max_attn_shift: 10
  coverage_penalty: 5.0
  ctc_weight: 0.4

beam_searcher_lm: !new:speechbrain.decoders.seq2seq.S2SRNNBeamSearchLM
  embedding: *id008
  decoder: *id009
  linear: *id010
  ctc_linear: *id013
  language_model: *id015
  bos_index: 0
  eos_index: 1
  blank_index: 2
  min_decode_ratio: 0
  max_decode_ratio: 1.0
  beam_size: 16
  eos_threshold: 10.0
  using_max_attn_shift: false
  max_attn_shift: 10
  coverage_penalty: 5.0
  ctc_weight: 0.4
  lm_weight: 0.5
  temperature: 1.25
  temperature_lm: 1.0


lr_annealing: &id018 !new:speechbrain.nnet.schedulers.NewBobScheduler
  initial_value: 0.002
  improvement_threshold: 0.0
  annealing_factor: 0.8
  patient: 0

homograph_extractor: !new:speechbrain.lobes.models.g2p.homograph.SubsequenceExtractor

seq_cost: &id016 !name:speechbrain.nnet.losses.nll_loss

  label_smoothing: 0.1

ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
  blank_index: 2

seq_cost_metric: &id017 !name:speechbrain.nnet.losses.nll_loss

  label_smoothing: 0.1
  reduction: batch

homograph_cost: !new:speechbrain.lobes.models.g2p.homograph.SubsequenceLoss
  seq_cost: *id016
seq_stats: !name:speechbrain.utils.metric_stats.MetricStats
  metric: *id017
seq_stats_homograph: !name:speechbrain.utils.metric_stats.MetricStats
  metric: *id017
classification_stats_homograph: !name:speechbrain.utils.metric_stats.ClassificationStats

per_stats: !name:speechbrain.utils.metric_stats.ErrorRateStats
per_stats_homograph: !name:speechbrain.utils.metric_stats.ErrorRateStats


model_output_keys:
- p_seq
- char_lens
- encoder_out

grapheme_encoder: &id027 !new:speechbrain.dataio.encoder.TextEncoder
phoneme_encoder: &id024 !new:speechbrain.dataio.encoder.TextEncoder


grapheme_tokenizer: !apply:speechbrain.lobes.models.g2p.dataio.lazy_init
  init: !name:speechbrain.tokenizers.SentencePiece.SentencePiece
    model_dir: grapheme_tokenizer
    bos_id: 0
    eos_id: 1
    unk_id: 2
    vocab_size: 512
    annotation_train: tokenizer_annotation_train.json
    annotation_read: char
    model_type: unigram                    # ["unigram", "bpe", "char"]
    character_coverage: 1.0
    annotation_format: json
    text_file: grapheme_annotations.txt

phoneme_tokenizer: &id022 !apply:speechbrain.lobes.models.g2p.dataio.lazy_init
  init: !name:speechbrain.tokenizers.SentencePiece.SentencePiece
    model_dir: phoneme_tokenizer
    bos_id: 0
    eos_id: 1
    unk_id: 2
    vocab_size: 512
    annotation_train: tokenizer_annotation_train.json
    annotation_read: phn
    model_type: unigram                   # ["unigram", "bpe", "char"]
    character_coverage: 1.0
    annotation_list_to_check: [tokenizer_annotation_valid.json]
    annotation_format: json
    text_file: phoneme_annotations.txt

out_phoneme_decoder_tok: &id025 !apply:speechbrain.lobes.models.g2p.dataio.char_map_detokenize
  tokenizer: *id022
  char_map: *id023
  token_space_index: 512
  wordwise: true

out_phoneme_decoder_raw: &id026 !name:speechbrain.lobes.models.g2p.dataio.text_decode

  encoder: *id024
out_phoneme_decoder: !apply:speechbrain.utils.hparams.choice
  value: false
  choices:
    true: *id025
    false: *id026
encode_pipeline:
  batch: false
  use_padded_data: true
  output_keys:
  - grapheme_list
  - grapheme_encoded_list
  - grapheme_encoded
  - word_emb
  init:
  - func: !name:speechbrain.lobes.models.g2p.dataio.enable_eos_bos
      encoder: *id027
      tokens: *id028
      bos_index: 0
      eos_index: 1
  - func: !name:speechbrain.lobes.models.g2p.dataio.enable_eos_bos
      encoder: *id024
      tokens: *id001
      bos_index: 0
      eos_index: 1
  steps:
  - func: !name:speechbrain.lobes.models.g2p.dataio.clean_pipeline
      graphemes: *id028
    takes: txt
    provides: txt_cleaned
  - func: !name:speechbrain.lobes.models.g2p.dataio.grapheme_pipeline
      grapheme_encoder: *id027
    takes: txt_cleaned
    provides:
    - grapheme_list
    - grapheme_encoded_list
    - grapheme_encoded_raw

  - func: !name:speechbrain.lobes.models.g2p.dataio.add_bos_eos
      encoder: *id027
    takes: grapheme_encoded_list
    provides:
    - grapheme_encoded
    - grapheme_len
    - grapheme_encoded_eos
    - grapheme_len_eos
  - func: !name:speechbrain.lobes.models.g2p.dataio.word_emb_pipeline
      word_emb: !ref <word_emb>
      grapheme_encoder: !ref <grapheme_encoder>
      use_word_emb: !ref <use_word_emb>
    takes:
    - txt
    - grapheme_encoded
    - grapheme_len
    provides: word_emb

decode_pipeline:
  batch: true
  output_keys:
  - phonemes
  steps:
  - func: !name:speechbrain.lobes.models.g2p.dataio.beam_search_pipeline
      beam_searcher: *id029
    takes:
    - char_lens
    - encoder_out
    provides:
    - hyps
    - scores
  - func: !apply:speechbrain.utils.hparams.choice
      value: false
      choices:
        true: !apply:speechbrain.lobes.models.g2p.dataio.char_map_detokenize
          tokenizer: *id022
          char_map: *id023
          token_space_index: 512
          wordwise: true
        false: !name:speechbrain.lobes.models.g2p.dataio.phoneme_decoder_pipeline
          phoneme_encoder: *id024
    takes:
    - hyps
    provides:
    - phonemes


pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
  loadables:
    model: *id014
    ctc_lin: *id013