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

Merge pull request #1019 from zh794390558/feat

Browse Source 
[bugfix] Kaldi Feature using dither in train
pull/1024/head
Jackwaterveg 3 years ago committed by GitHub
parent 7cddfd27f7 7a25ee26d9
commit 09931d2ccc
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 15 additions and 11 deletions
Show Stats Download Patch File Download Diff File

14
docs/source/released_model.md
Unescape View File

@ -4,13 +4,13 @@
### Acoustic Model Released in paddle 2.X
Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech | example link
:-------------:| :------------:| :-----: | -----: | :----------------- |:--------- | :---------- | :--------- | :-----------
[Ds2 Online Aishell S0 Model](https://deepspeech.bj.bcebos.com/release2.2/aishell/s0/ds2_online_aishll_CER8.02_release.tar.gz) | Aishell Dataset | Char-based | 345 MB | 2 Conv + 5 LSTM layers with only forward direction | 0.080218 |-| 151 h | [D2 Online Aishell S0 Example](../../examples/aishell/s0)
[Ds2 Offline Aishell S0 Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds2.offline.cer6p65.release.tar.gz)| Aishell Dataset | Char-based | 306 MB | 2 Conv + 3 bidirectional GRU layers| 0.065 |-| 151 h | [Ds2 Offline Aishell S0 Example](../../examples/aishell/s0)
[Conformer Online Aishell S1 Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.chunk.release.tar.gz) | Aishell Dataset | Char-based | 283 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention rescoring | 0.0594 |-| 151 h | [Conformer Online Aishell S1 Example](../../examples/aishell/s1)
[Conformer Offline Aishell S1 Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.release.tar.gz) | Aishell Dataset | Char-based | 284 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention rescoring | 0.0547 |-| 151 h | [Conformer Offline Aishell S1 Example](../../examples/aishell/s1)
[Conformer Librispeech S1 Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/conformer.release.tar.gz) | Librispeech Dataset | subword-based | 287 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention rescoring |-| 0.0325 | 960 h | [Conformer Librispeech S1 example](../../example/librispeech/s1)
[Transformer Librispeech S1 Model](https://deepspeech.bj.bcebos.com/release2.2/librispeech/s1/librispeech.s1.transformer.all.wer5p62.release.tar.gz) | Librispeech Dataset | subword-based | 131 MB | Encoder:Transformer, Decoder:Transformer, Decoding method: Attention rescoring |-| 0.0456 | 960 h | [Transformer Librispeech S1 example](../../example/librispeech/s1)
[Transformer Librispeech S2 Model](https://deepspeech.bj.bcebos.com/release2.2/librispeech/s2/libri_transformer_espnet_wer3p84.release.tar.gz) | Librispeech Dataset | subword-based | 131 MB | Encoder:Transformer, Decoder:Transformer, Decoding method: Attention |-| 0.0384 | 960 h | [Transformer Librispeech S2 example](../../example/librispeech/s2)
[Ds2 Online Aishell ASR0 Model](https://deepspeech.bj.bcebos.com/release2.2/aishell/s0/ds2_online_aishll_CER8.02_release.tar.gz) | Aishell Dataset | Char-based | 345 MB | 2 Conv + 5 LSTM layers with only forward direction | 0.080218 |-| 151 h | [D2 Online Aishell S0 Example](../../examples/aishell/asr0)
[Ds2 Offline Aishell ASR0 Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s0/aishell.s0.ds2.offline.cer6p65.release.tar.gz)| Aishell Dataset | Char-based | 306 MB | 2 Conv + 3 bidirectional GRU layers| 0.065 |-| 151 h | [Ds2 Offline Aishell S0 Example](../../examples/aishell/asr0)
[Conformer Online Aishell ASR1 Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.chunk.release.tar.gz) | Aishell Dataset | Char-based | 283 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention rescoring | 0.0594 |-| 151 h | [Conformer Online Aishell S1 Example](../../examples/aishell/s1)
[Conformer Offline Aishell ASR1 Model](https://deepspeech.bj.bcebos.com/release2.1/aishell/s1/aishell.release.tar.gz) | Aishell Dataset | Char-based | 284 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention rescoring | 0.0547 |-| 151 h | [Conformer Offline Aishell S1 Example](../../examples/aishell/s1)
[Conformer Librispeech ASR1 Model](https://deepspeech.bj.bcebos.com/release2.1/librispeech/s1/conformer.release.tar.gz) | Librispeech Dataset | subword-based | 287 MB | Encoder:Conformer, Decoder:Transformer, Decoding method: Attention rescoring |-| 0.0325 | 960 h | [Conformer Librispeech S1 example](../../example/librispeech/s1)
[Transformer Librispeech ASR1 Model](https://deepspeech.bj.bcebos.com/release2.2/librispeech/s1/librispeech.s1.transformer.all.wer5p62.release.tar.gz) | Librispeech Dataset | subword-based | 131 MB | Encoder:Transformer, Decoder:Transformer, Decoding method: Attention rescoring |-| 0.0456 | 960 h | [Transformer Librispeech S1 example](../../example/librispeech/s1)
[Transformer Librispeech ASR2 Model](https://deepspeech.bj.bcebos.com/release2.2/librispeech/s2/libri_transformer_espnet_wer3p84.release.tar.gz) | Librispeech Dataset | subword-based | 131 MB | Encoder:Transformer, Decoder:Transformer, Decoding method: Attention |-| 0.0384 | 960 h | [Transformer Librispeech S2 example](../../example/librispeech/s2)
### Acoustic Model Transformed from paddle 1.8
Acoustic Model | Training Data | Token-based | Size | Descriptions | CER | WER | Hours of speech

2
paddlespeech/s2t/frontend/featurizer/text_featurizer.py
Unescape View File

@ -56,6 +56,8 @@ class TextFeaturizer():
self.vocab_dict, self._id2token, self.vocab_list, self.unk_id, self.eos_id, self.blank_id = self._load_vocabulary_from_file(
vocab_filepath, maskctc)
self.vocab_size = len(self.vocab_list)
else:
logger.warning("TextFeaturizer: not have vocab file.")
if unit_type == 'spm':
spm_model = spm_model_prefix + '.model'

8
paddlespeech/s2t/transform/spectrogram.py
Unescape View File

@ -341,7 +341,7 @@ class LogMelSpectrogramKaldi():
self.eps = eps
self.remove_dc_offset = True
self.preemph = 0.97
self.dither = dither
self.dither = dither # only work in train mode
def __repr__(self):
return (
@ -361,11 +361,12 @@ class LogMelSpectrogramKaldi():
eps=self.eps,
dither=self.dither, ))
def __call__(self, x):
def __call__(self, x, train):
"""
Args:
x (np.ndarray): shape (Ti,)
train (bool): True, train mode.
Raises:
ValueError: not support (Ti, C)
@ -373,6 +374,7 @@ class LogMelSpectrogramKaldi():
Returns:
np.ndarray: (T, D)
"""
dither = self.dither if train else False
if x.ndim != 1:
raise ValueError("Not support x: [Time, Channel]")
@ -391,7 +393,7 @@ class LogMelSpectrogramKaldi():
nfft=self.n_fft,
lowfreq=self.fmin,
highfreq=self.fmax,
dither=self.dither,
dither=dither,
remove_dc_offset=self.remove_dc_offset,
preemph=self.preemph,
wintype=self.window)

Write Preview
Loading…
Cancel
Save