Skip to content

A No-Recurrence Sequence-to-Sequence Model for Speech Recognition

License

Notifications You must be signed in to change notification settings

ZhengkunTian/OpenTransformer

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

OpenTransformer

This is a speech-transformer model for end-to-end speech recognition. If you have any questions, please email to me. ([email protected])

Requirements

Pytorch >= 1.2.0 (<= 1.6.0)

Torchaudio >= 0.3.0

TO DO

  • Reduce the redundant computation by cache the previous states during inference

Function

  • Speech Transformer / Conformer

  • Label Smoothing

  • Tie Weights of Embedding with output softmax layer

  • Data Augmentation(SpecAugument)

  • Extract Fbank features in a online fashion

  • Read the feature with the kaldi or espnet format!

  • Visualization based Tensorboard [Will Be Updated Soon!]

  • Batch Beam Search with Length Penalty

  • Multiple Optimizers and Schedulers

  • Multiple Activation Functions in FFN

  • Multi GPU

  • LM Shollow Fusion

Prepare

vocab

# character idx
<PAD> 0
<S/E> 1
<UNK> 2
我 3
你 4
...

character

BAC009S0764W0139 国 家 统 计 局 的 数 据 显 示
BAC009S0764W0140 其 中 广 州 深 圳 甚 至 出 现 了 多 个 日 光 盘
BAC009S0764W0141 零 三 年 到 去 年
BAC009S0764W0142 市 场 基 数 已 不 可 同 日 而 语
BAC009S0764W0143 在 市 场 整 体 从 高 速 增 长 进 入 中 高 速 增 长 区 间 的 同 时
BAC009S0764W0144 一 线 城 市 在 价 格 较 高 的 基 础 上 整 体 回 升 并 领 涨 全 国
BAC009S0764W0145 绝 大 部 分 三 线 城 市 房 价 仍 然 下 降
BAC009S0764W0146 一 线 楼 市 成 交 量 激 增
BAC009S0764W0147 三 四 线 城 市 依 然 冷 清

if you want to compute features online, please make sure you have a wav.scp file.

# wav.scp
# id path
BAC009S0764W0139 /data/aishell/wav/BAC009S0764W0139.wav

Train

  • Single GPU
python run.py -c egs/aishell/conf/transformer.yaml
  • Multi GPU Training based DataParallel
python run.py -c egs/aishell/transformer.yaml -n 2 -g 0,1

Average the parameters of the last N epochs

python tools/average.py your_model_expdir 50 59    #   average the models from 50-th epoch to 59-th epoch

Eval

python eval.py -m model.pt

Experiments

Our Model can achieve a CER of 6.7% without CMVN, any external LM and joint-CTC training on AISHELL-1, which is better than 7.4% of Chain Model in Kaldi.

Acknowledge

OpenTransformer refer to ESPNET.

About

A No-Recurrence Sequence-to-Sequence Model for Speech Recognition

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •