alignment.yaml

# Generated 2026-02-02 from:
# /home/streamalign/streamASR/hparams/train_chunk_streaming.yaml
# yamllint disable
# ############################################################################
# Model: E2E ASR with Conformer Transducer (Character-level)
# Encoder: Conformer with Dynamic Chunk Training
# Decoder: LSTM + greedy/beam search
# Tokens: Character-level (no BPE)
# Losses: Transducer + CTC (optional)
# Training: Librispeech
# 
# Modified for character-level prediction and CER evaluation
# ############################################################################

seed: 3407
__set_seed: !apply:speechbrain.utils.seed_everything [3407]
output_folder: results/conformer_transducer_char/alignment
output_wer_folder: results/conformer_transducer_char/alignment/
save_folder: results/conformer_transducer_char/alignment/save
train_log: results/conformer_transducer_char/alignment/train_log.txt

# Data files
data_folder: /home/datasets/LibriTTS

# Files to hold the manifest data
train_json: !ref <save_folder>/train.json
valid_json: !ref <save_folder>/valid.json
test_json: !ref <save_folder>/test.json

# Data splits
train_split: ["train-clean-100", "train-clean-360", "train-other-500"]
valid_split: ["dev-clean", "dev-other"]
test_split: ["test-clean", "test-other"]

skip_prep: false
ckpt_interval_minutes: 5
####################### Training Parameters ####################################

number_of_epochs: 100
warmup_steps: 25000
num_workers: 4
batch_size_valid: 4
lr: 0.0008
weight_decay: 0.01
number_of_ctc_epochs: 60
ctc_weight: 0.3
ce_weight: 0.0
max_grad_norm: 5.0
loss_reduction: batchmean
precision: fp16

batch_size: 8
grad_accumulation_factor: 4
sorting: random
avg_checkpoints: 10

# Feature parameters
sample_rate: 16000
n_fft: 512
n_mels: 80
win_length: 32

# ============================================================================
# Streaming & Dynamic Chunk Training
# ============================================================================
streaming: true

dynchunktrain_config_sampler: !new:speechbrain.utils.dynamic_chunk_training.DynChunkTrainConfigRandomSampler
  chunkwise_prob: 0.6
  chunk_size_min: 8
  chunk_size_max: 32
  limited_left_context_prob: 0.75
  left_context_chunks_min: 2
  left_context_chunks_max: 32
   # Validation uses same streaming config as test
  valid_config: !new:speechbrain.utils.dynamic_chunk_training.DynChunkTrainConfig
    chunk_size: 16
    left_context_size: 8
   # Test config with specific streaming parameters
  test_config: !new:speechbrain.utils.dynamic_chunk_training.DynChunkTrainConfig
    chunk_size: 16
    left_context_size: 8

# Dataloader options
train_dataloader_opts:
  batch_size: 8
  num_workers: 4

valid_dataloader_opts:
  batch_size: 4

test_dataloader_opts:
  batch_size: 4

# Dynamic batching
dynamic_batching: true
max_batch_len: 150
max_batch_len_val: 50
num_bucket: 200

dynamic_batch_sampler:
  max_batch_len: 150
  max_batch_len_val: 50
  num_buckets: 200
  shuffle_ex: true
  batch_ordering: random
  max_batch_ex: 256

####################### Model Parameters #######################################

# Transformer/Conformer
d_model: 512
joint_dim: 640
nhead: 8
num_encoder_layers: 12
num_decoder_layers: 0
d_ffn: 2048
transformer_dropout: 0.1
activation: &id005 !name:torch.nn.GELU


# ============================================================================
# Character-level output
# 72 characters + blank = 73 tokens (CTC_VOCAB_SIZE)
# space/punct/a-z/accented (0-71) + blank (72)
# ============================================================================
output_neurons: 73

# Decoder
dec_dim: 512
dec_emb_dropout: 0.2
dec_dropout: 0.1
attention_type: RelPosMHAXL

# Decoding parameters
blank_index: 72
bos_index: 72
eos_index: 72
pad_index: 72
beam_size: 10
nbest: 1
state_beam: 2.3
expand_beam: 2.3

# If True uses torchaudio loss
use_torchaudio: false

epoch_counter: &id019 !new:speechbrain.utils.epoch_loop.EpochCounter

  limit: 100

normalize: &id013 !new:speechbrain.processing.features.InputNormalization
  norm_type: global
  update_until_epoch: 4

compute_features: &id021 !new:speechbrain.lobes.features.Fbank
  sample_rate: 16000
  n_fft: 512
  n_mels: 80
  win_length: 32

############################## Augmentations ###################################

speed_perturb: &id001 !new:speechbrain.augment.time_domain.SpeedPerturb
  orig_freq: 16000
  speeds: [95, 100, 105]

wav_augment: !new:speechbrain.augment.augmenter.Augmenter
  min_augmentations: 1
  max_augmentations: 1
  augment_prob: 1.0
  augmentations: [*id001]

time_drop: &id002 !new:speechbrain.augment.freq_domain.SpectrogramDrop
  drop_length_low: 12
  drop_length_high: 20
  drop_count_low: 5
  drop_count_high: 5
  replace: zeros

freq_drop: &id003 !new:speechbrain.augment.freq_domain.SpectrogramDrop
  drop_length_low: 20
  drop_length_high: 25
  drop_count_low: 2
  drop_count_high: 2
  replace: zeros
  dim: 2

time_warp: &id004 !new:speechbrain.augment.freq_domain.Warping

fea_augment: !new:speechbrain.augment.augmenter.Augmenter
  parallel_augment: false
  concat_original: false
  repeat_augment: 1
  shuffle_augmentations: false
  min_augmentations: 3
  max_augmentations: 3
  augment_prob: 1.0
  augmentations: [*id002, *id003, *id004]

############################## Pretrained ###################################
pretrained_source: speechbrain/asr-streaming-conformer-librispeech
pretrained_save: results/conformer_transducer_char/alignment/pretrained


############################## Models ##########################################

CNN: &id007 !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
  input_shape: (8, 10, 80)
  num_blocks: 2
  num_layers_per_block: 1
  out_channels: (64, 32)
  kernel_sizes: (3, 3)
  strides: (2, 2)
  residuals: (False, False)

Transformer: &id006 !new:models.TransformerASR.TransformerASR

  input_size: 640
  tgt_vocab: 73
  d_model: 512
  nhead: 8
  num_encoder_layers: 12
  num_decoder_layers: 0
  d_ffn: 2048
  dropout: 0.1
  activation: *id005
  encoder_module: conformer
  attention_type: RelPosMHAXL
  normalize_before: true
  output_hidden_states: true
  causal: false

enc: &id008 !new:models.TransformerASR.EncoderWrapper
  transformer: *id006
proj_ctc: &id014 !new:speechbrain.nnet.linear.Linear
  input_size: 640
  n_neurons: 73

proj_enc: &id016 !new:speechbrain.nnet.linear.Linear

  input_size: 512
  n_neurons: 640
  bias: false

proj_dec: &id015 !new:speechbrain.nnet.linear.Linear
  input_size: 512
  n_neurons: 640
  bias: false

ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
  blank_index: 72
  reduction: batchmean

# Character embedding (output_neurons - 1 because blank is handled separately)
emb: &id009 !new:speechbrain.nnet.embedding.Embedding
  num_embeddings: 73
  consider_as_one_hot: true
  blank_id: 72

dec: &id010 !new:speechbrain.nnet.RNN.LSTM
  input_shape: [null, null, 72]
  hidden_size: 512
  num_layers: 1
  re_init: true

ce_cost: !name:speechbrain.nnet.losses.nll_loss
  label_smoothing: 0.1

Tjoint: &id011 !new:speechbrain.nnet.transducer.transducer_joint.Transducer_joint
  joint: sum
  nonlinearity: *id005
transducer_lin: &id012 !new:speechbrain.nnet.linear.Linear
  input_size: 640
  n_neurons: 73
  bias: false

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

transducer_cost: !name:speechbrain.nnet.losses.transducer_loss
  blank_index: 72
  use_torchaudio: false

modules:
  CNN: *id007
  enc: *id008
  emb: *id009
  dec: *id010
  Tjoint: *id011
  transducer_lin: *id012
  normalize: *id013
  proj_ctc: *id014
  proj_dec: *id015
  proj_enc: *id016
model: &id017 !new:torch.nn.ModuleList
- [*id007, *id008, *id009, *id010, *id016, *id015, *id014, *id012]
Greedysearcher: &id020 !new:speechbrain.decoders.transducer.TransducerBeamSearcher
  decode_network_lst: [*id009, *id010, *id015]
  tjoint: *id011
  classifier_network: [*id012]
  blank_id: 72
  beam_size: 1
  nbest: 1

Beamsearcher: !new:speechbrain.decoders.transducer.TransducerBeamSearcher
  decode_network_lst: [*id009, *id010, *id015]
  tjoint: *id011
  classifier_network: [*id012]
  blank_id: 72
  beam_size: 10
  nbest: 1
  state_beam: 2.3
  expand_beam: 2.3

opt_class: !name:torch.optim.AdamW
  lr: 0.0008
  betas: (0.9, 0.98)
  eps: 1.e-8
  weight_decay: 0.01

noam_annealing: &id018 !new:speechbrain.nnet.schedulers.NoamScheduler
  lr_initial: 0.0008
  n_warmup_steps: 25000


############################## Logging and Checkpointing #######################

checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
  checkpoints_dir: results/conformer_transducer_char/alignment/save
  recoverables:
    model: *id017
    scheduler: *id018
    normalizer: *id013
    counter: *id019
train_logger: !new:speechbrain.utils.train_logger.WandBLogger
  initializer: !name:wandb.init
    project: streamingAlign
    name: subalign
    dir: results/conformer_transducer_char/alignment/wandb
    reinit: true
    resume: false

# CER computer (character-level)
cer_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
  split_tokens: true

############################## Streaming Inference #############################

# Streaming configuration for inference
# chunk_size: frames per chunk (8 frames ≈ 320ms after 4x subsampling)
# left_context_size: number of left context chunks
streaming_chunk_size: 16
streaming_left_context: 8

# Create streaming context for the decoder
make_decoder_streaming_context: !name:speechbrain.decoders.transducer.TransducerGreedySearcherStreamingContext

# Streaming decoding function
decoding_function: !name:speechbrain.decoders.transducer.TransducerBeamSearcher.transducer_greedy_decode_streaming
- *id020
fea_streaming_extractor: !new:speechbrain.lobes.features.StreamingFeatureWrapper
  module: !new:speechbrain.nnet.containers.LengthsCapableSequential
  - *id021
  - *id013
  - *id007
  properties: !apply:speechbrain.utils.filter_analysis.stack_filter_properties
  - [*id021, *id007]