Reference: https://arxiv.org/abs/2005.08100
speech_config:
sample_rate: 16000
frame_ms: 25
stride_ms: 10
feature_type: log_mel_spectrogram
num_feature_bins: 80
preemphasis: 0.97
normalize_signal: True
normalize_feature: True
normalize_per_feature: False
decoder_config:
vocabulary: null
target_vocab_size: 1024
max_subword_length: 4
blank_at_zero: True
beam_width: 5
norm_score: True
model_config:
name: conformer
subsampling:
type: conv2
kernel_size: 3
strides: 2
filters: 144
positional_encoding: sinusoid_concat
dmodel: 144
num_blocks: 16
head_size: 36
num_heads: 4
mha_type: relmha
kernel_size: 32
fc_factor: 0.5
dropout: 0.1
embed_dim: 320
embed_dropout: 0.0
num_rnns: 1
rnn_units: 320
rnn_type: lstm
layer_norm: True
joint_dim: 320
learning_config:
augmentations:
after:
time_masking:
num_masks: 10
mask_factor: 100
p_upperbound: 0.2
freq_masking:
num_masks: 1
mask_factor: 27
dataset_config:
train_paths: ...
eval_paths: ...
test_paths: ...
tfrecords_dir: ...
optimizer_config:
warmup_steps: 10000
beta1: 0.9
beta2: 0.98
epsilon: 1e-9
running_config:
batch_size: 4
num_epochs: 22
outdir: ...
log_interval_steps: 400
save_interval_steps: 400
eval_interval_steps: 1000Training, see python examples/conformer/train_conformer.py --help
Testing, see python examples/conformer/train_conformer.py --help
TFLite Conversion, see python examples/conformer/tflite_conformer.py --help
Summary
- Number of subwords: 1031
- Maxium length of a subword: 4
- Subwords corpus: all training sets, dev sets and test-clean
- Number of parameters: 10,341,639
- Positional Encoding Type: sinusoid concatenation
Pretrained and Config, go to drive
Transducer Loss
Error Rates
| Test-clean | WER (%) | CER (%) |
|---|---|---|
| Greedy | 6.4476862 | 2.51828337 |