You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
PaddleSpeech/examples/aishell3/ernie_sat/conf/default.yaml

287 lines
4.2 KiB

# This configuration tested on 8 GPUs (A100) with 80GB GPU memory.
# It takes around 3 days to finish the training,You can adjust
# batch_size、num_workers here and ngpu in local/train.sh for your machine
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 24000 # sr
n_fft: 2048 # FFT size (samples).
n_shift: 300 # Hop size (samples). 12.5ms
win_length: 1200 # Window length (samples). 50ms
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
# Only used for feats_type != raw
fmin: 80 # Minimum frequency of Mel basis.
fmax: 7600 # Maximum frequency of Mel basis.
n_mels: 80 # The number of mel basis.
mean_phn_span: 8
mlm_prob: 0.8
###########################################################
# DATA SETTING #
###########################################################
batch_size: 40
num_workers: 8
###########################################################
# MODEL SETTING #
###########################################################
model:
text_masking: false
postnet_layers: 5
postnet_filts: 5
postnet_chans: 256
encoder_type: conformer
decoder_type: conformer
enc_input_layer: sega_mlm
enc_pre_speech_layer: 0
enc_cnn_module_kernel: 7
enc_attention_dim: 384
enc_attention_heads: 2
enc_linear_units: 1536
enc_num_blocks: 4
enc_dropout_rate: 0.2
enc_positional_dropout_rate: 0.2
enc_attention_dropout_rate: 0.2
enc_normalize_before: true
enc_macaron_style: true
enc_use_cnn_module: true
enc_selfattention_layer_type: legacy_rel_selfattn
enc_activation_type: swish
enc_pos_enc_layer_type: legacy_rel_pos
enc_positionwise_layer_type: conv1d
enc_positionwise_conv_kernel_size: 3
dec_cnn_module_kernel: 31
dec_attention_dim: 384
dec_attention_heads: 2
dec_linear_units: 1536
dec_num_blocks: 4
dec_dropout_rate: 0.2
dec_positional_dropout_rate: 0.2
dec_attention_dropout_rate: 0.2
dec_macaron_style: true
dec_use_cnn_module: true
dec_selfattention_layer_type: legacy_rel_selfattn
dec_activation_type: swish
dec_pos_enc_layer_type: legacy_rel_pos
dec_positionwise_layer_type: conv1d
dec_positionwise_conv_kernel_size: 3
###########################################################
# OPTIMIZER SETTING #
###########################################################
scheduler_params:
d_model: 384
warmup_steps: 4000
grad_clip: 1.0
###########################################################
# TRAINING SETTING #
###########################################################
max_epoch: 1500
num_snapshots: 50
###########################################################
# OTHER SETTING #
###########################################################
seed: 0
token_list:
- <blank>
- <unk>
- d
- sp
- sh
- ii
- j
- zh
- l
- x
- b
- g
- uu
- e5
- h
- q
- m
- i1
- t
- z
- ch
- f
- s
- u4
- ix4
- i4
- n
- i3
- iu3
- vv
- ian4
- ix2
- r
- e4
- ai4
- k
- ing2
- a1
- en2
- ui4
- ong1
- uo3
- u2
- u3
- ao4
- ee
- p
- an1
- eng2
- i2
- in1
- c
- ai2
- ian2
- e2
- an4
- ing4
- v4
- ai3
- a5
- ian3
- eng1
- ong4
- ang4
- ian1
- ing1
- iy4
- ao3
- ang1
- uo4
- u1
- iao4
- iu4
- a4
- van2
- ie4
- ang2
- ou4
- iang4
- ix1
- er4
- iy1
- e1
- en1
- ui2
- an3
- ei4
- ong2
- uo1
- ou3
- uo2
- iao1
- ou1
- an2
- uan4
- ia4
- ia1
- ang3
- v3
- iu2
- iao3
- in4
- a3
- ei3
- iang3
- v2
- eng4
- en3
- aa
- uan1
- v1
- ao1
- ve4
- ie3
- ai1
- ing3
- iang1
- a2
- ui1
- en4
- en5
- in3
- uan3
- e3
- ie1
- ve2
- ei2
- in2
- ix3
- uan2
- iang2
- ie2
- ua4
- ou2
- uai4
- er2
- eng3
- uang3
- un1
- ong3
- uang4
- vn4
- un2
- iy3
- iz4
- ui3
- iao2
- iong4
- un4
- van4
- ao2
- uang1
- iy5
- o2
- ei1
- ua1
- iu1
- uang2
- er5
- o1
- un3
- vn1
- vn2
- o4
- ve1
- van3
- ua2
- er3
- iong3
- van1
- ia2
- iy2
- ia3
- iong1
- uo5
- oo
- ve3
- ou5
- uai3
- ian5
- iong2
- uai2
- uai1
- ua3
- vn3
- ia5
- ie5
- ueng1
- o5
- o3
- iang5
- ei5
- <sos/eos>