Efficient Conformer v1 for non-streaming ASR
Specification: https://github.com/wenet-e2e/wenet/pull/1636
Aishell-1 Results
- Feature info:
- using fbank feature, cmvn, speed perturb, dither
- Training info:
- train_u2++_efficonformer_v1.yaml
- 8 gpu, batch size 16, acc_grad 1, 200 epochs
- lr 0.001, warmup_steps 25000
- Model info:
- Model Params: 48,488,347
- Downsample rate: 1/4 (conv2d) * 1/2 (efficonformer block)
- encoder_dim 256, output_size 256, head 8, linear_units 2048
- num_blocks 12, cnn_module_kernel 15, group_size 3
- Decoding info:
- ctc_weight 0.5, reverse_weight 0.3, average_num 20
decoding mode | full | 18 | 16 |
---|---|---|---|
attention decoder | 4.99 | 5.13 | 5.16 |
ctc prefix beam search | 4.98 | 5.23 | 5.23 |
attention rescoring | 4.64 | 4.86 | 4.85 |
Start to Use
Install WeNet follow: https://wenet.org.cn/wenet/install.html#install-for-training
Decode
cd wenet/examples/aishell/s0
dir=exp/wenet_efficient_conformer_aishell_v1/
ctc_weight=0.5
reverse_weight=0.3
decoding_chunk_size=-1
mode="attention_rescoring"
test_dir=$dir/test_${mode}
mkdir -p $test_dir
# Decode
nohup python wenet/bin/recognize.py --gpu 0 \
--mode $mode \
--config $dir/train.yaml \
--data_type "raw" \
--test_data data/test/data.list \
--checkpoint $dir/final.pt \
--beam_size 10 \
--batch_size 1 \
--penalty 0.0 \
--dict $dir/words.txt \
--ctc_weight $ctc_weight \
--reverse_weight $reverse_weight \
--result_file $test_dir/text \
${decoding_chunk_size:+--decoding_chunk_size $decoding_chunk_size} > logs/decode_aishell.log &
# CER
python tools/compute-cer.py --char=1 --v=1 \
data/test/text $test_dir/text > $test_dir/cer.txt