---
# System prepended metadata

title: Week 1C - 翻譯蒟蒻
tags: ['2023', Penut]

---

# 翻譯蒟蒻
## Week 1C

----

## 簡介

最近 Meta 釋出 SeamlessM4T
支援 100 種語言，語音與文字互相翻譯
包含文字翻譯、語音辨識、語音合成等


[Blog](https://ai.meta.com/blog/seamless-m4t/) | [GitHub](https://github.com/facebookresearch/seamless_communication) | [Paper](https://dl.fbaipublicfiles.com/seamless/seamless_m4t_paper.pdf)

Demo
[Official](https://seamless.metademolab.com/) | [HF Space](https://huggingface.co/spaces/facebook/seamless_m4t) | [Colab](https://colab.research.google.com/drive/1r4IQAiVAtEz-4yIvqoEcA1lFZLbXFPTV?hl=en) | [Colab Gradio](https://colab.research.google.com/drive/110d7rjVyAQX7Gg9ixtRD3pk0nP-tM0Iy?hl=en)

----

## 架構

![](https://i.imgur.com/HiriCMF.png)

----

## 環境

```bash=
git clone https://github.com/facebookresearch/seamless_communication.git
cd seamless_communication
pip install .
conda install -y -c conda-forge libsndfile
```

----

## 初始化

```python=
import torch
from seamless_communication.models.inference import Translator

model = "seamlessM4T_large"  # or "seamlessM4T_medium"

translator = Translator(
    model,
    vocoder_name_or_card="vocoder_36langs",
    device=torch.device("cuda:0"),
    dtype=torch.float16,
)
```

模型大小約 10 GB，在 Colab 上只要花一分鐘下載

----

## 翻譯文字

```python=
results = translator.predict(
    input="今天天氣真好！",
    task_str="T2ST",
    src_lang="eng",
    tgt_lang="cmn",
    spkr=-1,
)
translated_text, wav, sr = results
```

`spkr` 可以選不同 TTS 語者
`sr` 為輸出音檔用的 Sample Rate
可用的語言可以參考[這份檔案](https://github.com/facebookresearch/seamless_communication/blob/main/src/seamless_communication/assets/cards/vocoder_36langs.yaml)

----

## 輸出音檔

透過 `torchaudio` 輸出 `.wav` 音檔

```python=
import torchaudio

torchaudio.save("output.wav", wav.squeeze(0).cpu(), sr)
```

在 Jupyter 上直接顯示播放元件

```python=
from IPython.display import Audio

Audio(wav.squeeze(0).cpu(), rate=sr)
```

----

## 翻譯音檔

```python=
results = translator.predict(
    input="output.wav",
    task_str="S2ST",
    tgt_lang="cmn",
)
translated_text, wav, sr = results
```

---

# 其他分享

----

## 秒殺容器

有些容器在停止時會花很久的時間
是因為 `docker stop` 會先送出 `SIGTERM`
十秒後才送出 `SYSKILL` 的訊號

可以直接 `docker stop -t 1` 秒殺

[ChatGPT Reference](https://chat.openai.com/share/15f53a67-39eb-41da-8d94-0a2ed533464b)

----

## LLM & Image

最近 Bard 與 Bing Chat 都支援上傳圖片了

[Bard with Image](https://bard.google.com/?hl=en)

----

## Python 不朽

Python 預計加入 Immortal Objects
提升記憶體效率 (Share Memory)
解除平行運算限制

[News Link](https://www.ithome.com.tw/news/158384)

----

## 遙遙領先的 IDE

來自中國、領先世界的 CEC-IDE

[GitHub Issue](https://github.com/microsoft/vscode/issues/191229)

----

## Shell Script Trap

Trap 機制可以捕捉類似 Ctrl + C 的訊號
讓 Shell Script 可以做一些例外處理或垃圾回收

```bash=
cleanup() {
  rm -f /tmp/tempfile
}
trap cleanup EXIT
touch /tmp/tempfile
```

ChatGPT Reference - [Link 1](https://chat.openai.com/share/4a6ec6e8-65b2-4249-85ff-3bf8911b17d6) | [Link 2](https://chat.openai.com/share/61aa4120-d7be-4cb2-ad9e-6cf8ac13e21d)

----

## TTS

[TTS Maker](https://ttsmaker.com/zh-hk)
[雅婷文字轉語音](https://tts.yating.tw/#exp)
[雅婷逐字稿](https://asr.yating.tw/)

----

## OpenCC Converter

[自製簡繁轉換器](https://penut85420-opencc-converter.hf.space/)
[OpenCC GitHub](https://github.com/BYVoid/OpenCC)