msb-public
/
PaddleSpeech
mirror of https://github.com/PaddlePaddle/PaddleSpeech
1
0
Code Issues Projects Releases Wiki Activity

[asr] update librispeech conformer result (#1116)

Browse Source 
* update librispeech result

* change conf order
pull/1117/head
Hui Zhang 3 years ago committed by GitHub
parent 65f684806e
commit c4a79ccea4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 105 additions and 126 deletions
Show Stats Download Patch File Download Diff File

13
examples/librispeech/asr1/RESULTS.md
Unescape View File

@ -1,15 +1,18 @@
# LibriSpeech
## Conformer
train: Epoch 70, 4 V100-32G, best avg: 20
| Model | Params | Config | Augmentation| Test set | Decode method | Loss | WER |
| --- | --- | --- | --- | --- | --- | --- | --- |
| conformer | 47.63 M | conf/conformer.yaml | spec_aug + shift | test-clean | attention | 6.738649845123291 | 0.041159 |
| conformer | 47.63 M | conf/conformer.yaml | spec_aug + shift | test-clean | ctc_greedy_search | 6.738649845123291 | 0.039847 |
| conformer | 47.63 M | conf/conformer.yaml | spec_aug + shift | test-clean | ctc_prefix_beam_search | 6.738649845123291 | 0.039790 |
| conformer | 47.63 M | conf/conformer.yaml | spec_aug + shift | test-clean | attention_rescoring | 6.738649845123291 | 0.034617 |
| conformer | 47.63 M | conf/conformer.yaml | spec_aug | test-clean | attention | 6.433612394332886 | 0.039771 |
| conformer | 47.63 M | conf/conformer.yaml | spec_aug | test-clean | ctc_greedy_search | 6.433612394332886 | 0.040342 |
| conformer | 47.63 M | conf/conformer.yaml | spec_aug | test-clean | ctc_prefix_beam_search | 6.433612394332886 | 0.040342 |
| conformer | 47.63 M | conf/conformer.yaml | spec_aug | test-clean | attention_rescoring | 6.433612394332886 | 0.033761 |
## Chunk Conformer
| Model | Params | Config | Augmentation| Test set | Decode method | Chunk Size & Left Chunks | Loss | WER |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| conformer | 47.63 M | conf/chunk_conformer.yaml | spec_aug + shift | test-clean | attention | 16, -1 | 7.11 | 0.063193 |
@ -20,7 +23,7 @@
## Transformer
train: Epoch 120, 4 V100-32G, 27 Day, avg: 10
train: Epoch 120, 4 V100-32G, 27 Day, best avg: 10
| Model | Params | Config | Augmentation| Test set | Decode method | Loss | WER |
| --- | --- | --- | --- | --- | --- | --- | --- |

71
examples/librispeech/asr1/conf/chunk_conformer.yaml
Unescape View File

@ -1,41 +1,3 @@
# https://yaml.org/type/float.html
data:
train_manifest: data/manifest.train
dev_manifest: data/manifest.dev
test_manifest: data/manifest.test
min_input_len: 0.5
max_input_len: 30.0
min_output_len: 0.0
max_output_len: 400.0
min_output_input_ratio: 0.05
max_output_input_ratio: 100.0
collator:
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml
batch_size: 16
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
# network architecture
model:
cmvn_file:
@ -80,6 +42,39 @@ model:
length_normalized_loss: false
data:
train_manifest: data/manifest.train
dev_manifest: data/manifest.dev
test_manifest: data/manifest.test
collator:
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml
batch_size: 16
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
training:
n_epoch: 240
accum_grad: 8

69
examples/librispeech/asr1/conf/chunk_transformer.yaml
Unescape View File

@ -1,41 +1,3 @@
# https://yaml.org/type/float.html
data:
train_manifest: data/manifest.train
dev_manifest: data/manifest.dev
test_manifest: data/manifest.test
min_input_len: 0.5 # second
max_input_len: 30.0 # second
min_output_len: 0.0 # tokens
max_output_len: 400.0 # tokens
min_output_input_ratio: 0.05
max_output_input_ratio: 100.0
collator:
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml
batch_size: 64
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
# network architecture
model:
cmvn_file:
@ -73,6 +35,37 @@ model:
length_normalized_loss: false
data:
train_manifest: data/manifest.train
dev_manifest: data/manifest.dev
test_manifest: data/manifest.test
collator:
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml
batch_size: 64
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
training:
n_epoch: 120
accum_grad: 1

78
examples/librispeech/asr1/conf/conformer.yaml
Unescape View File

@ -1,41 +1,3 @@
# https://yaml.org/type/float.html
data:
train_manifest: data/manifest.train
dev_manifest: data/manifest.dev
test_manifest: data/manifest.test-clean
min_input_len: 0.5 # seconds
max_input_len: 30.0 # seconds
min_output_len: 0.0 # tokens
max_output_len: 400.0 # tokens
min_output_input_ratio: 0.05
max_output_input_ratio: 100.0
collator:
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml
batch_size: 16
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
# network architecture
model:
cmvn_file:
@ -76,8 +38,40 @@ model:
length_normalized_loss: false
data:
train_manifest: data/manifest.train
dev_manifest: data/manifest.dev
test_manifest: data/manifest.test-clean
collator:
vocab_filepath: data/lang_char/vocab.txt
unit_type: 'spm'
spm_model_prefix: 'data/lang_char/bpe_unigram_5000'
mean_std_filepath: ""
augmentation_config: conf/preprocess.yaml
batch_size: 16
raw_wav: True # use raw_wav or kaldi feature
spectrum_type: fbank #linear, mfcc, fbank
feat_dim: 80
delta_delta: False
dither: 1.0
target_sample_rate: 16000
max_freq: None
n_fft: None
stride_ms: 10.0
window_ms: 25.0
use_dB_normalization: True
target_dB: -20
random_seed: 0
keep_transcription_text: False
sortagrad: True
shuffle_method: batch_shuffle
num_workers: 2
training:
n_epoch: 120
n_epoch: 70
accum_grad: 8
global_grad_clip: 3.0
optim: adam
@ -98,13 +92,7 @@ decoding:
batch_size: 64
error_rate_type: wer
decoding_method: attention # 'attention', 'ctc_greedy_search', 'ctc_prefix_beam_search', 'attention_rescoring'
lang_model_path: data/lm/common_crawl_00.prune01111.trie.klm
alpha: 2.5
beta: 0.3
beam_size: 10
cutoff_prob: 1.0
cutoff_top_n: 0
num_proc_bsearch: 8
ctc_weight: 0.5 # ctc weight for attention rescoring decode mode.
decoding_chunk_size: -1 # decoding chunk size. Defaults to -1.
# <0: for decoding, use full chunk.

Write Preview
Loading…
Cancel
Save