config

Config File Documentation

Description of parameters available in config file are listed here, checkout example configs for how to use.

asr_model

Optimizer

Parameter	Description	Note
type	The name of pytorch optimizer for training.	Tested: Adam/Adadelta
learning_rate	Learning rate for optimizer.
joint_ctc	Weight on CTC loss function for multi-task learning, see Joint CTC training for detail. Set it to 0 for attention-based seq2seq (e.g. LAS) learning with cross entropy only.	Pure CTC (set this parameter to 1.0) is not available currently.

Encoder

Parameter	Description	Note
enc_type	Encoder architecture. For VGGBiRNN, 4x input time subsampling will be performed with CNN.	Available now:VGGBiRNN/BiRNN/RNN
sample_rate	Sample rate for each RNN layer, concatenated with _. For each layer, the length of output on the time dimension will be input/sample_rate.
sample_style	The down sampling mechanism. concat will concatenate multiple time steps according to sample rate into one vector, drop will drop the unsampled timesteps.	Available now:concat/drop
dim	Number of cells for each RNN layer (per direction), concatenated with _.	Depth must match sample_rate
dropout	Dropout between each layer, concatenated with _.	Depth must match sample_rate
rnn_cell	RNN cell of all layer.	Tested: LSTM

Attention

Parameter	Description	Note
att_mode	Attention mechanism option. dot is the vanilla attention and loc indicates the location-based attention.	Available now: dot/loc
dim	The dimension of all networks in attention (if there are any).
proj	Apply an additional transform layer to encoder feature before calculating energy.
num_head	Under development...	Under development...

clm

If enabled, adversarial training between ASR and CLM will be imposed. This would allow ASR to learn from additional text source. See the paper for more detailed description.

solver

Parameter	Description	Note
dataset	Dataset name for experiment.	Available now: timit/librispeech
data_path	File path of the pre-processed dataset.
n_jobs	Number of workers for Pytorch Dataloader.
max_timestep	Upper limit of the length of input, training data with acoustic feature longer than the limit will be dropped. (0 = No restriction)	Only applies to training data.
max_label_len	Upper limit of the length of output, training data with label sequence longer than the limit will be dropped. (0 = No restriction)	Only applies to training data.
train_set	List containing the splits to be used as training data. Check the name of csv files under data_path	e.g. ['train-clean-100','train-clean-360'] when dataset=librispeech
batch_size	Batch size for training.
apex	Enable faster optimization, see APEX for more details.	Install apex first before setting this option to True.
total_steps	Total number of steps to train.
tf_start	Apply Scheduled Sampling with teacher forcing rate at the beginning of training. With teacher forcing rate set to 1, ASR will never sample input from its own output.
tf_end	Teacher forcing rate at the end of training (linearly decay during training, 1 for no scheduled sampling at all).
dev_set	List containing the splits to be used as validation data.	See train_set for how to use.
dev_batch_size	Batch size for validation.
dev_step	Interval between each validation step (i.e. validation will be performed every <dev_step> step).
test_set	List containing the splits to be decoded in the testing phase.	See train_set for how to use.
max_decode_step_ratio	The maximum decoding time step will be max_decode_step_ratioxinput length during testing (in case beam search might never end). Decoding will stop if is predicted or max decoding step reached.
decode_beam_size	Beam size for beam search algorithm.
decode_ctc_weight	The weight on CTC prefix probability during decoding, see joint CTC decoding for more detail.
decode_lm_weight	The weight on probability predicted by RNNLM during decoding, see joint RNNLM decoding for more detail.
decode_lm_path	The path of pre-trained RNNLM.