Update README.md
Browse files
README.md
CHANGED
@@ -3,13 +3,15 @@ license: apache-2.0
|
|
3 |
---
|
4 |
## Chinese-English ASR model using k2-zipformer-streaming
|
5 |
### AIShell-1 and Wenetspeech testset results with modified-beam-search streaming decode using epoch-14.pt
|
6 |
-
| AIShell-1 | TEST_NET | TEST_MEETING |
|
7 |
-
|
8 |
-
| 3.19 | 9.58 | 9.51 ||
|
|
|
9 |
|
10 |
-
### Training
|
11 |
```
|
12 |
nohup ./pruned_transducer_stateless7_streaming/train.py --world-size 8 --num-epochs 30 --start-epoch 1 --feedforward-dims "1024,1024,1536,1536,1024" --exp-dir pruned_transducer_stateless7_streaming/exp --max-duration 360 > pruned_transducer_stateless7_streaming/exp/nohup.zipformer &
|
|
|
13 |
```
|
14 |
|
15 |
### Model unit is char+bpe as `data/lang_char_bpe/tokens.txt`
|
@@ -31,4 +33,4 @@ dims': '192,192,192,192,192', 'encoder_unmasked_dims': '256,256,256,256,256', 'z
|
|
31 |
_dim': 512, 'joiner_dim': 512, 'short_chunk_size': 50, 'num_left_chunks': 4, 'decode_chunk_len': 32, 'manifest_dir': PosixPath('data/fbank'), 'max_duration': 360, 'bucketing
|
32 |
_sampler': True, 'num_buckets': 300, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': True, 'return_cuts': True, 'num_wor
|
33 |
kers': 8, 'enable_spec_aug': True, 'spec_aug_time_warp_factor': 80, 'enable_musan': True, 'training_subset': '12k_hour', 'blank_id': 0, 'vocab_size': 6254}
|
34 |
-
```
|
|
|
3 |
---
|
4 |
## Chinese-English ASR model using k2-zipformer-streaming
|
5 |
### AIShell-1 and Wenetspeech testset results with modified-beam-search streaming decode using epoch-14.pt
|
6 |
+
| decode_chunk_len | AIShell-1 | TEST_NET | TEST_MEETING |
|
7 |
+
|------------------|-----------|----------|--------------|
|
8 |
+
| 32 | 3.19 | 9.58 | 9.51 ||
|
9 |
+
| 64 | 3.04 | 8.97 | 8.83 ||
|
10 |
|
11 |
+
### Training and decoding commands
|
12 |
```
|
13 |
nohup ./pruned_transducer_stateless7_streaming/train.py --world-size 8 --num-epochs 30 --start-epoch 1 --feedforward-dims "1024,1024,1536,1536,1024" --exp-dir pruned_transducer_stateless7_streaming/exp --max-duration 360 > pruned_transducer_stateless7_streaming/exp/nohup.zipformer &
|
14 |
+
nohup ./pruned_transducer_stateless7_streaming/decode.py --epoch 6 --avg 1 --exp-dir ./pruned_transducer_stateless7_streaming/exp --max-duration 600 --decode-chunk-len 32 --decoding-method modified_beam_search --beam-size 4 > nohup.zipformer.deocode &
|
15 |
```
|
16 |
|
17 |
### Model unit is char+bpe as `data/lang_char_bpe/tokens.txt`
|
|
|
33 |
_dim': 512, 'joiner_dim': 512, 'short_chunk_size': 50, 'num_left_chunks': 4, 'decode_chunk_len': 32, 'manifest_dir': PosixPath('data/fbank'), 'max_duration': 360, 'bucketing
|
34 |
_sampler': True, 'num_buckets': 300, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': True, 'return_cuts': True, 'num_wor
|
35 |
kers': 8, 'enable_spec_aug': True, 'spec_aug_time_warp_factor': 80, 'enable_musan': True, 'training_subset': '12k_hour', 'blank_id': 0, 'vocab_size': 6254}
|
36 |
+
```
|