# It contains the default values for training an LSTM-Transducer ASR model, large size (~170M for bidirectional and ~130M for unidirectional) with Transducer loss and sub-word encoding. # Architecture and training config: # Default learning parameters in this config are set for effective batch size of 2K. To train it with smaller effective # batch sizes, you may need to re-tune the learning parameters or use higher accumulate_grad_batches. # Followed the architecture suggested in the following paper: # 'STREAMING END-TO-END SPEECH RECOGNITION FOR MOBILE DEVICES' by Yanzhang He et al. (https://arxiv.org/pdf/1811.06621.pdf) # You may find more info about LSTM-Transducer here: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#lstm-transducer # Pre-trained models of LSTM-Transducer can be found here: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/results.html name: "LSTM-Transducer-BPE" model: sample_rate: 16000 compute_eval_loss: false # eval samples can be very long and exhaust memory. Disable computation of transducer loss during validation/testing with this flag. log_prediction: true # enables logging sample predictions in the output during training skip_nan_grad: false model_defaults: enc_hidden: 640 pred_hidden: 640 joint_hidden: 640 rnn_hidden_size: 2048 train_ds: manifest_filepath: ??? sample_rate: ${model.sample_rate} batch_size: 16 # you may increase batch_size if your memory allows shuffle: true num_workers: 4 pin_memory: true use_start_end_token: false trim_silence: false max_duration: 16.7 # it is set for LibriSpeech, you may need to update it for your dataset min_duration: 0.1 # tarred datasets is_tarred: false tarred_audio_filepaths: null shuffle_n: 2048 # bucketing params bucketing_strategy: "synced_randomized" bucketing_batch_size: null validation_ds: manifest_filepath: ??? sample_rate: ${model.sample_rate} batch_size: 16 shuffle: false num_workers: 4 pin_memory: true use_start_end_token: false test_ds: manifest_filepath: null sample_rate: ${model.sample_rate} batch_size: 16 shuffle: false num_workers: 4 pin_memory: true use_start_end_token: false # You may find more detail on how to train a tokenizer at: /scripts/tokenizers/process_asr_text_tokenizer.py tokenizer: dir: ??? # path to directory which contains either tokenizer.model (bpe) or vocab.txt (for wpe) type: bpe # Can be either bpe (SentencePiece tokenizer) or wpe (WordPiece tokenizer) preprocessor: _target_: nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor sample_rate: ${model.sample_rate} normalize: "per_feature" window_size: 0.025 window_stride: 0.01 window: "hann" features: 80 n_fft: 512 frame_splicing: 1 dither: 0.00001 pad_to: 0 spec_augment: _target_: nemo.collections.asr.modules.SpectrogramAugmentation freq_masks: 2 # set to zero to disable it time_masks: 10 # set to zero to disable it freq_width: 27 time_width: 0.05 encoder: _target_: nemo.collections.asr.modules.RNNEncoder feat_in: ${model.preprocessor.features} n_layers: 8 d_model: 2048 proj_size: ${model.model_defaults.pred_hidden} # you may set it if you need different output size other than the default d_model rnn_type: "lstm" # it can be lstm, gru or rnn bidirectional: true # need to set it to false if you want to make the model causal # Sub-sampling params subsampling: stacking # stacking, vggnet or striding subsampling_factor: 4 subsampling_conv_channels: -1 # set to -1 to make it equal to the d_model ### regularization dropout: 0.2 # The dropout used in most of the Conformer Modules decoder: _target_: nemo.collections.asr.modules.RNNTDecoder normalization_mode: null # Currently only null is supported for export. random_state_sampling: false # Random state sampling: https://arxiv.org/pdf/1910.11455.pdf blank_as_pad: true # This flag must be set in order to support exporting of RNNT models + efficient inference. prednet: pred_hidden: ${model.model_defaults.pred_hidden} pred_rnn_layers: 2 t_max: null dropout: 0.2 rnn_hidden_size: 2048 joint: _target_: nemo.collections.asr.modules.RNNTJoint log_softmax: null # 'null' would set it automatically according to CPU/GPU device preserve_memory: false # dramatically slows down training, but might preserve some memory # Fuses the computation of prediction net + joint net + loss + WER calculation # to be run on sub-batches of size `fused_batch_size`. # When this flag is set to true, consider the `batch_size` of *_ds to be just `encoder` batch size. # `fused_batch_size` is the actual batch size of the prediction net, joint net and transducer loss. # Using small values here will preserve a lot of memory during training, but will make training slower as well. # An optimal ratio of fused_batch_size : *_ds.batch_size is 1:1. # However, to preserve memory, this ratio can be 1:8 or even 1:16. # Extreme case of 1:B (i.e. fused_batch_size=1) should be avoided as training speed would be very slow. fuse_loss_wer: true fused_batch_size: 16 jointnet: joint_hidden: ${model.model_defaults.joint_hidden} activation: "relu" dropout: 0.2 decoding: strategy: "greedy_batch" # can be greedy, greedy_batch, beam, tsd, alsd. # greedy strategy config greedy: max_symbols: 10 # beam strategy config beam: beam_size: 2 return_best_hypothesis: False score_norm: true tsd_max_sym_exp: 50 # for Time Synchronous Decoding alsd_max_target_len: 2.0 # for Alignment-Length Synchronous Decoding loss: loss_name: "default" warprnnt_numba_kwargs: # FastEmit regularization: https://arxiv.org/abs/2010.11148 # You may enable FastEmit to reduce the latency of the model for streaming # using fastemit_lambda=1e-3 can help the accuracy of the model when it is unidirectional fastemit_lambda: 0.0 # Recommended values to be in range [1e-4, 1e-2], 0.001 is a good start. # Adds Gaussian noise to the gradients of the decoder to avoid overfitting variational_noise: start_step: 0 std: 0.0 optim: name: adamw lr: 5.0 # optimizer arguments betas: [0.9, 0.98] weight_decay: 1e-2 # scheduler setup sched: name: NoamAnnealing d_model: ${model.encoder.d_model} # scheduler config override warmup_steps: 10000 warmup_ratio: null min_lr: 1e-6 trainer: devices: -1 # number of GPUs, -1 would use all available GPUs num_nodes: 1 max_epochs: 500 max_steps: -1 # computed at runtime if not set val_check_interval: 1.0 # Set to 0.25 to check 4 times per epoch, or an int for number of iterations accelerator: auto strategy: ddp accumulate_grad_batches: 1 gradient_clip_val: 0.3 precision: 32 # Should be set to 16 for O1 and O2 to enable the AMP. log_every_n_steps: 10 # Interval of logging. enable_progress_bar: True resume_from_checkpoint: null # The path to a checkpoint file to continue the training, restores the whole state including the epoch, step, LR schedulers, apex, etc. num_sanity_val_steps: 0 # number of steps to perform validation steps for sanity check the validation process before starting the training, setting to 0 disables it check_val_every_n_epoch: 1 # number of evaluations on validation every n epochs sync_batchnorm: true enable_checkpointing: False # Provided by exp_manager logger: false # Provided by exp_manager benchmark: false # needs to be false for models with variable-length speech input as it slows down training exp_manager: exp_dir: null name: ${name} create_tensorboard_logger: true create_checkpoint_callback: true checkpoint_callback_params: # in case of multiple validation sets, first one is used monitor: "val_wer" mode: "min" save_top_k: 5 always_save_nemo: True # saves the checkpoints as nemo files instead of PTL checkpoints # you need to set these two to True to continue the training resume_if_exists: false resume_ignore_no_checkpoint: false # You may use this section to create a W&B logger create_wandb_logger: false wandb_logger_kwargs: name: null project: null