File size: 3,196 Bytes
664aefc 564731a 664aefc 564731a 3774ffb 564731a 3774ffb 564731a 3774ffb b882505 3774ffb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 |
---
license: apache-2.0
datasets:
- tedlium3
language:
- en
metrics:
- wer
---
### TedLium3 Zipformer
**`rnnt_type=regular`**
The WERs are
| | dev | test | comment |
|------------------------------------|------------|------------|------------------------------------------|
| greedy search | 6.74 | 6.16 | --epoch 50, --avg 22, --max-duration 500 |
| beam search (beam size 4) | 6.56 | 5.95 | --epoch 50, --avg 22, --max-duration 500 |
| modified beam search (beam size 4) | 6.54 | 6.00 | --epoch 50, --avg 22, --max-duration 500 |
| fast beam search (set as default) | 6.91 | 6.28 | --epoch 50, --avg 22, --max-duration 500 |
The training command for reproducing is given below:
```
export CUDA_VISIBLE_DEVICES="0,1,2,3"
./zipformer/train.py \
--use-fp16 true \
--world-size 4 \
--num-epochs 50 \
--start-epoch 0 \
--exp-dir zipformer/exp \
--max-duration 1000
```
The tensorboard training log can be found at
https://tensorboard.dev/experiment/AKXbJha0S9aXyfmuvG4h5A/#scalars
The decoding command is:
```
epoch=50
avg=22
## greedy search
./zipformer/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir zipformer/exp \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 500
## beam search
./zipformer/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir zipformer/exp \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 500 \
--decoding-method beam_search \
--beam-size 4
## modified beam search
./zipformer/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir zipformer/exp \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 500 \
--decoding-method modified_beam_search \
--beam-size 4
## fast beam search
./zipformer/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir ./zipformer/exp \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 1500 \
--decoding-method fast_beam_search \
--beam 4 \
--max-contexts 4 \
--max-states 8
```
**`rnnt_type=modified`**
Using the codes from this PR https://github.com/k2-fsa/icefall/pull/1125.
The WERs are
| | dev | test | comment |
|------------------------------------|------------|------------|------------------------------------------|
| greedy search | 6.32 | 5.83 | --epoch 50, --avg 22, --max-duration 500 |
| modified beam search (beam size 4) | 6.16 | 5.79 | --epoch 50, --avg 22, --max-duration 500 |
| fast beam search (set as default) | 6.30 | 5.89 | --epoch 50, --avg 22, --max-duration 500 |
The training command for reproducing is given below:
```
export CUDA_VISIBLE_DEVICES="0,1,2,3"
./zipformer/train.py \
--use-fp16 true \
--world-size 4 \
--num-epochs 50 \
--start-epoch 0 \
--exp-dir zipformer/exp \
--max-duration 1000 \
--rnnt-type modified
```
The tensorboard training log can be found at
https://tensorboard.dev/experiment/3d4bYmbJTGiWQQaW88CVEQ/#scalars
The decoding commands are same as above. |