wsj0_2mix_skim_small_causal开源语音增强模型 - 免费部署实现混合语音信号分离

首页

Wsj0 2mix Skim Small Causal

由 lichenda 开发

这是一个基于ESPnet框架训练的语音增强模型，专门用于处理wsj0_2mix数据集中的混合语音信号分离任务。

音频增强英语#实时语音增强 #因果性模型 #低延迟处理

下载量 26

发布时间 : 5/17/2023

模型简介

该模型采用skim架构，具有因果性处理能力，适用于实时语音增强场景，能够有效分离混合语音中的不同说话人信号。

模型特点

因果处理能力

模型采用因果结构设计，适合实时语音处理应用场景

轻量级架构

小型skim架构设计，在保持性能的同时降低计算复杂度

多说话人分离

能够有效分离混合语音中的两个说话人信号

模型能力

语音增强

说话人分离

实时语音处理

使用案例

语音通信

会议语音增强

在多人会议场景中分离不同说话人的声音

STOI指标达到94.20，SDR指标14.33

语音识别预处理

ASR前端处理

为语音识别系统提供更干净的输入信号

可提升语音识别系统在嘈杂环境中的准确率

🚀 ESPnet2 ENH模型

本模型是一个音频增强模型，基于ESPnet框架训练，使用wsj0_2mix数据集，可用于音频到音频的处理任务。

🚀 快速开始

安装ESPnet

若你还未安装ESPnet，请按照ESPnet安装说明进行操作。

运行示例

cd espnet
git checkout 3897ed8380bfb526d5c5dd2197eccbffbba7d8f8
pip install -e .
cd egs2/wsj0_2mix/enh1
./run.sh --skip_data_prep false --skip_train true --download_model lichenda/wsj0_2mix_skim_small_causal

📚 详细文档

模型训练信息

此模型由李辰达（Chenda Li）使用espnet中的wsj0_2mix配方进行训练。

实验结果

环境信息

日期: Wed May 10 20:30:26 CST 2023
Python版本: 3.9.16 (main, Mar 8 2023, 14:00:05) [GCC 11.2.0]
ESPnet版本: espnet 202304
PyTorch版本: pytorch 2.0.1
Git哈希值: 3897ed8380bfb526d5c5dd2197eccbffbba7d8f8
- 提交日期: Tue May 9 13:27:37 2023 +0800

评估指标

数据集	STOI	SAR	SDR	SIR	SI_SNR
enhanced_cv_min_8k	93.41	15.54	14.92	24.87	14.51
enhanced_tt_min_8k	94.20	15.00	14.33	24.18	13.92

增强配置

展开

config: conf/tuning/train_enh_skim_causal_small.yaml
print_config: false
log_level: INFO
dry_run: false
iterator_type: chunk
output_dir: exp/enh_train_enh_skim_causal_small_raw
ngpu: 1
seed: 0
num_workers: 4
num_att_plot: 3
dist_backend: nccl
dist_init_method: env://
dist_world_size: null
dist_rank: null
local_rank: 0
dist_master_addr: null
dist_master_port: null
dist_launcher: null
multiprocessing_distributed: false
unused_parameters: false
sharded_ddp: false
cudnn_enabled: true
cudnn_benchmark: false
cudnn_deterministic: true
collect_stats: false
write_collected_feats: false
max_epoch: 150
patience: 50
val_scheduler_criterion:
- valid
- loss
early_stopping_criterion:
- valid
- loss
- min
best_model_criterion:
-   - valid
    - si_snr_loss
    - min
-   - valid
    - loss
    - min
keep_nbest_models: 1
nbest_averaging_interval: 0
grad_clip: 5.0
grad_clip_type: 2.0
grad_noise: false
accum_grad: 1
no_forward_run: false
resume: true
train_dtype: float32
use_amp: false
log_interval: null
use_matplotlib: true
use_tensorboard: true
create_graph_in_tensorboard: false
use_wandb: false
wandb_project: null
wandb_id: null
wandb_entity: null
wandb_name: null
wandb_model_log_interval: -1
detect_anomaly: false
pretrain_path: null
init_param: []
ignore_init_mismatch: false
freeze_param: []
num_iters_per_epoch: null
batch_size: 16
valid_batch_size: null
batch_bins: 1000000
valid_batch_bins: null
train_shape_file:
- exp/enh_stats_8k/train/speech_mix_shape
- exp/enh_stats_8k/train/speech_ref1_shape
- exp/enh_stats_8k/train/speech_ref2_shape
valid_shape_file:
- exp/enh_stats_8k/valid/speech_mix_shape
- exp/enh_stats_8k/valid/speech_ref1_shape
- exp/enh_stats_8k/valid/speech_ref2_shape
batch_type: folded
valid_batch_type: null
fold_length:
- 80000
- 80000
- 80000
sort_in_batch: descending
sort_batch: descending
multiple_iterator: false
chunk_length: 32000,16000,8000
chunk_shift_ratio: 0.5
num_cache_chunks: 1024
chunk_excluded_key_prefixes: []
train_data_path_and_name_and_type:
-   - dump/raw/tr_min_8k/wav.scp
    - speech_mix
    - sound
-   - dump/raw/tr_min_8k/spk1.scp
    - speech_ref1
    - sound
-   - dump/raw/tr_min_8k/spk2.scp
    - speech_ref2
    - sound
valid_data_path_and_name_and_type:
-   - dump/raw/cv_min_8k/wav.scp
    - speech_mix
    - sound
-   - dump/raw/cv_min_8k/spk1.scp
    - speech_ref1
    - sound
-   - dump/raw/cv_min_8k/spk2.scp
    - speech_ref2
    - sound
allow_variable_data_keys: false
max_cache_size: 0.0
max_cache_fd: 32
valid_max_cache_size: null
exclude_weight_decay: false
exclude_weight_decay_conf: {}
optim: adam
optim_conf:
    lr: 0.001
    eps: 1.0e-08
    weight_decay: 0
scheduler: steplr
scheduler_conf:
    step_size: 2
    gamma: 0.97
init: xavier_uniform
model_conf:
    stft_consistency: false
    loss_type: mask_mse
    mask_type: null
criterions:
-   name: si_snr
    conf: {}
    wrapper: pit
    wrapper_conf:
        weight: 1.0
        independent_perm: true
speech_volume_normalize: null
rir_scp: null
rir_apply_prob: 1.0
noise_scp: null
noise_apply_prob: 1.0
noise_db_range: '13_15'
short_noise_thres: 0.5
use_reverberant_ref: false
num_spk: 1
num_noise_type: 1
sample_rate: 8000
force_single_channel: false
dynamic_mixing: false
utt2spk: null
dynamic_mixing_gain_db: 0.0
encoder: conv
encoder_conf:
    channel: 128
    kernel_size: 8
    stride: 4
separator: skim
separator_conf:
    causal: true
    num_spk: 2
    layer: 3
    nonlinear: relu
    unit: 384
    segment_size: 50
    dropout: 0.0
    mem_type: hc
    seg_overlap: false
decoder: conv
decoder_conf:
    channel: 128
    kernel_size: 8
    stride: 4
mask_module: multi_mask
mask_module_conf: {}
preprocessor: null
preprocessor_conf: {}
required:
- output_dir
version: '202304'
distributed: false

引用ESPnet

如果你在研究中使用了ESPnet，请引用以下文献：

@inproceedings{watanabe2018espnet,
  author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
  title={{ESPnet}: End-to-End Speech Processing Toolkit},
  year={2018},
  booktitle={Proceedings of Interspeech},
  pages={2207--2211},
  doi={10.21437/Interspeech.2018-1456},
  url={http://dx.doi.org/10.21437/Interspeech.2018-1456}
}


@inproceedings{ESPnet-SE,
  author = {Chenda Li and Jing Shi and Wangyou Zhang and Aswin Shanmugam Subramanian and Xuankai Chang and
  Naoyuki Kamo and Moto Hira and Tomoki Hayashi and Christoph B{"{o}}ddeker and Zhuo Chen and Shinji Watanabe},
  title = {ESPnet-SE: End-To-End Speech Enhancement and Separation Toolkit Designed for {ASR} Integration},
  booktitle = {{IEEE} Spoken Language Technology Workshop, {SLT} 2021, Shenzhen, China, January 19-22, 2021},
  pages = {785--792},
  publisher = {{IEEE}},
  year = {2021},
  url = {https://doi.org/10.1109/SLT48900.2021.9383615},
  doi = {10.1109/SLT48900.2021.9383615},
  timestamp = {Mon, 12 Apr 2021 17:08:59 +0200},
  biburl = {https://dblp.org/rec/conf/slt/Li0ZSCKHHBC021.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}


@inproceedings{liSkimSkippingMemory2022,
  title = {Skim: {{Skipping Memory Lstm}} for {{Low-Latency Real-Time Continuous Speech Separation}}},
  shorttitle = {Skim},
  booktitle = {{{ICASSP}} 2022 - 2022 {{IEEE International Conference}} on {{Acoustics}}, {{Speech}} and {{Signal Processing}} ({{ICASSP}})},
  author = {Li, Chenda and Yang, Lei and Wang, Weiqin and Qian, Yanmin},
  year = {2022},
  month = may,
  pages = {681--685},
  issn = {2379-190X},
  doi = {10.1109/ICASSP43922.2022.9746372},
}

或者引用arXiv上的文献：

@misc{watanabe2018espnet,
  title={ESPnet: End-to-End Speech Processing Toolkit},
  author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
  year={2018},
  eprint={1804.00015},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}