SeamlessM4T 本地部署 + 语音翻译APP 【干货】

SeamlessM4T

SeamlessM4T是Meta发布的一款能够转录和翻译近100种语言的一体化AI大模型。可以到官方演示地址进行体验。可以直接使用我改好的项目。

本地部署

一、拉取项目

# 指定项目地址
cd /opt
# 拉取项目
git clone https://github.com/facebookresearch/seamless_communication.git
1
2
3
4

二、安装依赖

# 创建虚拟环境
mkvirtualenv -p python3 seamless
# 安装依赖
cd /opt/seamless_communication
pip install .
1
2
3
4
5

三、获取模型

在项目的seamless_communication\src\seamless_communication\assets\cards文件路径下

seamlessM4T_large.yaml
seamlessM4T_medium.yaml
unity_nllb-100.yaml
unity_nllb-200.yaml
vocoder_36langs.yaml
xlsr2_1b_v2.yaml
1
2
3
4
5
6

写有对应的模型资源地址：

checkpoint: "https://huggingface.co/facebook/seamless-m4t-large/resolve/main/multitask_unity_large.pt"
checkpoint: "https://huggingface.co/facebook/seamless-m4t-medium/resolve/main/multitask_unity_medium.pt"
tokenizer: "https://huggingface.co/facebook/seamless-m4t-large/resolve/main/tokenizer.model"
tokenizer: "https://huggingface.co/facebook/seamless-m4t-medium/resolve/main/tokenizer.model"
checkpoint: "https://huggingface.co/facebook/seamless-m4t-vocoder/resolve/main/vocoder_36langs.pt"
checkpoint: "https://dl.fbaipublicfiles.com/seamlessM4T/models/unit_extraction/xlsr2_1b_v2.pt"
1
2
3
4
5
6

将对应的模型资源下载到本地对应位置后，修改配置内的地址

checkpoint: "file://opt/seamless_communication/models/large/multitask_unity_large.pt"
checkpoint: "file://opt/seamless_communication/models/medium/multitask_unity_medium.pt"
tokenizer: "file://opt/seamless_communication/models/large/tokenizer.model"
tokenizer: "file://opt/seamless_communication/models/medium/tokenizer.model"
checkpoint: "file://opt/seamless_communication/models/vocoder/vocoder_36langs.pt"
checkpoint: "file://opt/seamless_communication/models/xlsr2/xlsr2_1b_v2.pt"
1
2
3
4
5
6

四、修改依赖

将官方demo文件夹内的依赖文件/seamless_communication/demo/requirements.txt中的

git+https://github.com/facebookresearch/seamless_communication
1

修改为使用本地安装（因为更换了模型资源地址为本地）

/opt/seamless_communication
1

再执行依赖安装

cd /opt/seamless_communication/demo
workon seamless
pip install -r requirements.txt
1
2
3

如果事先执行过使用远程依赖安装的，则需要先卸载远程版本依赖

pip uninstall seamless_communication
1

再安装本地版本依赖

pip install /opt/seamless_communication
1

五、异常处理

官方demo中有一段代码是用于获取远程音频资源，在国内由于网络问题可能无法获取

# Download sample input audio files
filenames = ["assets/sample_input.mp3", "assets/sample_input_2.mp3"]
for filename in filenames:
    hf_hub_download(
        repo_id="facebook/seamless_m4t",
        repo_type="space",
        filename=filename,
        local_dir=".",
    )
1
2
3
4
5
6
7
8
9

可以手动在官方案例下载后，将sample_input.mp3、sample_input_2.mp3放到对应目录/opt/seamless_communication/demo/assets
并注释掉如上代码片段

六、运行Demo

workon seamless
cd /opt/seamless_communication/demo
python app.py
1
2
3

就可以看到运行成功

Running on local URL:  http://127.0.0.1:7860
1

6.1 修改启动端口

# 将
# demo.queue().launch()
# 修改成
demo.queue().launch(share=True, inbrowser=True, server_name='127.0.0.1', server_port=8080)
1
2
3
4

6.2 Gradio无法构建可访问链接

下载文件：

wget https://cdn-media.huggingface.co/frpc-gradio-0.2/frpc_linux_amd64
1

重命名下载的文件：

mv frpc_linux_amd64 frpc_linux_amd64_v0.2
1

将文件移动到指定位置：

mv frpc_linux_amd64_v0.2 /root/.virtualenvs/glm/lib/python3.8/site-packages/gradio
1

找到文件

cd /root/.virtualenvs/glm/lib/python3.8/site-packages/gradio
1

授予权限

chmod 777 frpc_linux_amd64_v0.2
1

6.3 异常处理

Traceback (most recent call last):
  File "app.py", line 587, in <module>
    input_audio_mic = gr.Audio(
  File "/root/.virtualenvs/seamless/lib/python3.8/site-packages/gradio/component_meta.py", line 152, in wrapper
    return fn(self, **kwargs)
TypeError: __init__() got an unexpected keyword argument 'source'
1
2
3
4
5
6

注释掉

# source="microphone",
# source="upload",
1
2

七、使用Flask自定义接口

在demo同级目录下，创建有Flask接口文件seamless_communication\demo\api.py，跳转到对应目录后，执行python api.py即可

7.1 特殊说明

因为生成的语音文件，默认保存到本地服务器的指定文件夹内，需要使用Nginx配置将该文件夹设置允许访问

vi /etc/nginx/nginx.conf
1

http {
    # 加入如下内容
	server {
		listen       8081;
		server_name  localhost;
		location /media/ {
			alias /opt/seamless_communication/media/out/;
			autoindex on;
		}
	}


}
1
2
3
4
5
6
7
8
9
10
11
12
13

service nginx restart
sudo chmod -R 755 /opt/seamless_communication/media/out/
1
2

八、 Flutter App

简单开发了一个配合Flask接口使用的FlutterAPP，可以自行根据项目代码编译。