mirror of
https://github.com/modelscope/FunASR
synced 2025-09-15 14:48:36 +08:00
3360a1d945
* boundary aware transducer * resolve conflict * delete type check * add bat egs and results --------- Co-authored-by: aky15 <ankeyu.aky@11.17.44.249>
109 lines
1.9 KiB
YAML
109 lines
1.9 KiB
YAML
encoder: chunk_conformer
|
|
encoder_conf:
|
|
activation_type: swish
|
|
positional_dropout_rate: 0.5
|
|
time_reduction_factor: 2
|
|
embed_vgg_like: false
|
|
subsampling_factor: 4
|
|
linear_units: 2048
|
|
output_size: 512
|
|
attention_heads: 8
|
|
dropout_rate: 0.5
|
|
positional_dropout_rate: 0.5
|
|
attention_dropout_rate: 0.5
|
|
cnn_module_kernel: 15
|
|
num_blocks: 12
|
|
|
|
# decoder related
|
|
rnnt_decoder: rnnt
|
|
rnnt_decoder_conf:
|
|
embed_size: 512
|
|
hidden_size: 512
|
|
embed_dropout_rate: 0.5
|
|
dropout_rate: 0.5
|
|
use_embed_mask: true
|
|
|
|
predictor: bat_predictor
|
|
predictor_conf:
|
|
idim: 512
|
|
threshold: 1.0
|
|
l_order: 1
|
|
r_order: 1
|
|
return_accum: true
|
|
|
|
joint_network_conf:
|
|
joint_space_size: 512
|
|
|
|
# frontend related
|
|
frontend: wav_frontend
|
|
frontend_conf:
|
|
fs: 16000
|
|
window: hamming
|
|
n_mels: 80
|
|
frame_length: 25
|
|
frame_shift: 10
|
|
lfr_m: 1
|
|
lfr_n: 1
|
|
|
|
|
|
# Auxiliary CTC
|
|
model: bat
|
|
model_conf:
|
|
auxiliary_ctc_weight: 0.0
|
|
cif_weight: 1.0
|
|
r_d: 3
|
|
r_u: 5
|
|
|
|
# minibatch related
|
|
use_amp: true
|
|
|
|
# optimization related
|
|
accum_grad: 1
|
|
grad_clip: 5
|
|
max_epoch: 100
|
|
val_scheduler_criterion:
|
|
- valid
|
|
- loss
|
|
best_model_criterion:
|
|
- - valid
|
|
- cer_transducer
|
|
- min
|
|
keep_nbest_models: 10
|
|
|
|
optim: adam
|
|
optim_conf:
|
|
lr: 0.001
|
|
scheduler: warmuplr
|
|
scheduler_conf:
|
|
warmup_steps: 25000
|
|
|
|
specaug: specaug
|
|
specaug_conf:
|
|
apply_time_warp: true
|
|
time_warp_window: 5
|
|
time_warp_mode: bicubic
|
|
apply_freq_mask: true
|
|
freq_mask_width_range:
|
|
- 0
|
|
- 40
|
|
num_freq_mask: 2
|
|
apply_time_mask: true
|
|
time_mask_width_range:
|
|
- 0
|
|
- 50
|
|
num_time_mask: 5
|
|
|
|
dataset_conf:
|
|
data_names: speech,text
|
|
data_types: sound,text
|
|
shuffle: True
|
|
shuffle_conf:
|
|
shuffle_size: 2048
|
|
sort_size: 500
|
|
batch_conf:
|
|
batch_type: token
|
|
batch_size: 25000
|
|
num_workers: 8
|
|
|
|
log_interval: 50
|