孫燕姿果然不愧是孫燕姿，不愧為南洋理工大學(xué)的高材生，近日她在個(gè)人官方媒體博客上寫了一篇英文版的長(zhǎng)文，正式回應(yīng)現(xiàn)在滿城風(fēng)雨的“AI孫燕姿”現(xiàn)象，流行天后展示了超人一等的智識(shí)水平，行文優(yōu)美，綿恒雋永，對(duì)AIGC藝術(shù)表現(xiàn)得極其克制，又相當(dāng)寬容，充滿了語(yǔ)言上的古典之美，表現(xiàn)出了“任彼如泰山壓頂，我只當(dāng)清風(fēng)拂面”的博大胸懷。

本次我們利用edge-tts和Sadtalker庫(kù)讓AI孫燕姿朗誦本尊的博文，讓流行天后念給你聽。

之前我們?cè)?jīng)使用百度開源的PaddleGAN視覺效果模型中一個(gè)子模塊Wav2lip實(shí)現(xiàn)了人物口型與輸入的歌詞語(yǔ)音同步，但Wav2lip的問(wèn)題是虛擬人物的動(dòng)態(tài)效果只能局限在嘴唇附近，事實(shí)上，音頻和不同面部動(dòng)作之間的連接是不同的，也就是說(shuō)，雖然嘴唇運(yùn)動(dòng)與音頻的聯(lián)系最強(qiáng)，但可以通過(guò)不同的頭部姿勢(shì)和眨眼來(lái)反作用于音頻。

【資料圖】

和Wav2lip相比，SadTaker是一種通過(guò)隱式3D系數(shù)調(diào)制的風(fēng)格化音頻驅(qū)動(dòng)Talking頭部視頻生成的庫(kù)，一方面，它從音頻中生成逼真的運(yùn)動(dòng)系數(shù)（例如，頭部姿勢(shì)、嘴唇運(yùn)動(dòng)和眨眼），并單獨(dú)學(xué)習(xí)每個(gè)運(yùn)動(dòng)以減少不確定性。對(duì)于表達(dá)，通過(guò)從的僅嘴唇運(yùn)動(dòng)系數(shù)和重建的渲染三維人臉上的感知損失（唇讀損失，面部landmark loss）中提取系數(shù)，設(shè)計(jì)了一種新的音頻到表達(dá)系數(shù)網(wǎng)絡(luò)。

對(duì)于程序化的頭部姿勢(shì)，通過(guò)學(xué)習(xí)給定姿勢(shì)的殘差，使用條件VAE來(lái)對(duì)多樣性和逼真的頭部運(yùn)動(dòng)進(jìn)行建模。在生成逼真的3DMM系數(shù)后，通過(guò)一種新穎的3D感知人臉渲染來(lái)驅(qū)動(dòng)源圖像。并且通過(guò)源和驅(qū)動(dòng)的無(wú)監(jiān)督3D關(guān)鍵點(diǎn)生成扭曲場(chǎng)，并扭曲參考圖像以生成最終視頻。

Sadtalker可以單獨(dú)配置，也可以作為Stable-Diffusion-Webui的插件而存在，這里推薦使用Stable-Diffusion插件的形式，因?yàn)檫@樣Stable-Diffusion和Sadtalker可以共用一套WebUI的界面，更方便將Stable-Diffusion生成的圖片做成動(dòng)態(tài)效果。

進(jìn)入到Stable-Diffusion的項(xiàng)目目錄：

cd stable-diffusion-webui

啟動(dòng)服務(wù)：

python3.10 webui.py

程序返回：

Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr  5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]  Version: v1.3.0  Commit hash: 20ae71faa8ef035c31aa3a410b707d792c8203a3  Installing requirements  Launching Web UI with arguments: --xformers --opt-sdp-attention --api --lowvram  Loading weights [b4d453442a] from D:\work\stable-diffusion-webui\models\Stable-diffusion\protogenV22Anime_protogenV22.safetensors  load Sadtalker Checkpoints from D:\work\stable-diffusion-webui\extensions\SadTalker\checkpoints  Creating model from config: D:\work\stable-diffusion-webui\configs\v1-inference.yaml  LatentDiffusion: Running in eps-prediction mode  DiffusionWrapper has 859.52 M params.  Running on local URL:  http://127.0.0.1:7860

代表啟動(dòng)成功，隨后http://localhost:7860

選擇插件(Extensions)選項(xiàng)卡

點(diǎn)擊從url安裝，輸入插件地址：github.com/Winfredy/SadTalker

安裝成功后，重啟WebUI界面。

接著需要手動(dòng)下載相關(guān)的模型文件：

https://pan.baidu.com/s/1nXuVNd0exUl37ISwWqbFGA?pwd=sadt

隨后將模型文件放入項(xiàng)目的stable-diffusion-webui/extensions/SadTalker/checkpoints/目錄即可。

接著配置一下模型目錄的環(huán)境變量：

set SADTALKER_CHECKPOINTS=D:/stable-diffusion-webui/extensions/SadTalker/checkpoints/

至此，SadTalker就配置好了。

之前的歌曲復(fù)刻是通過(guò)So-vits庫(kù)對(duì)原歌曲的音色進(jìn)行替換和預(yù)測(cè)，也就是說(shuō)需要原版的歌曲作為基礎(chǔ)數(shù)據(jù)。但目前的場(chǎng)景顯然有別于歌曲替換，我們首先需要將文本轉(zhuǎn)換為語(yǔ)音，才能替換音色。

這里使用edge-tts庫(kù)進(jìn)行文本轉(zhuǎn)語(yǔ)音操作：

import asyncio    import edge_tts    TEXT = """    As my AI voice takes on a life of its own while I despair over my overhanging stomach and my children"s every damn thing, I can"t help but want to write something about it.    My fans have officially switched sides and accepted that I am indeed 冷門歌手 while my AI persona is the current hot property. I mean really, how do you fight with someone who is putting out new albums in the time span of minutes.    Whether it is ChatGPT or AI or whatever name you want to call it, this "thing" is now capable of mimicking and/or conjuring,  unique and complicated content by processing a gazillion chunks of information while piecing and putting together in a most coherent manner the task being asked at hand. Wait a minute, isn"t that what humans do? The very task that we have always convinced ourselves; that the formation of thought or opinion is not replicable by robots, the very idea that this is beyond their league, is now the looming thing that will threaten thousands of human conjured jobs. Legal, medical, accountancy, and currently, singing a song.     You will protest, well I can tell the difference, there is no emotion or variance in tone/breath or whatever technical jargon you can come up with. Sorry to say, I suspect that this would be a very short term response.    Ironically, in no time at all, no human will be able to rise above that. No human will be able to have access to this amount of information AND make the right calls OR make the right mistakes (ok mayyyybe I"m jumping ahead). This new technology will be able to churn out what exactly EVERYTHING EVERYONE  needs. As indie or as warped or as psychotic as you can get, there"s probably a unique content that could be created just for you. You are not special you are already predictable and also unfortunately malleable.    At this point, I feel like a popcorn eater with the best seat in the theatre. (Sidenote: Quite possibly in this case no tech is able to predict what it"s like to be me, except when this is published then ok it"s free for all). It"s like watching that movie that changed alot of our lives Everything Everywhere All At Once, except in this case, I don"t think it will be the idea of love that will save the day.     In this boundless sea of existence, where anything is possible, where nothing matters, I think it will be purity of thought, that being exactly who you are will be enough.     With this I fare thee well.    """    VOICE = "en-HK-YanNeural"  OUTPUT_FILE = "./test_en1.mp3"      async def _main() -> None:      communicate = edge_tts.Communicate(TEXT, VOICE)      await communicate.save(OUTPUT_FILE)      if __name__ == "__main__":      asyncio.run(_main())

音頻使用英文版本的女聲：en-HK-YanNeural，關(guān)于edge-tts，請(qǐng)移步：口播神器,基于Edge,微軟TTS(text-to-speech)文字轉(zhuǎn)語(yǔ)音免費(fèi)開源庫(kù)edge-tts語(yǔ)音合成實(shí)踐(Python3.10)，這里不再贅述。

隨后再將音頻文件的音色替換為AI孫燕姿的音色即可：AI天后,在線飆歌,人工智能AI孫燕姿模型應(yīng)用實(shí)踐，復(fù)刻《遙遠(yuǎn)的歌》，原唱晴子(Python3.10)。

準(zhǔn)備好生成的圖片以及音頻文件后，就可以在本地進(jìn)行推理操作了，訪問(wèn) localhost:7860

這里輸入?yún)?shù)選擇full，如此會(huì)保留整個(gè)圖片區(qū)域，否則只保留頭部部分。

生成效果：

SadTalker會(huì)根據(jù)音頻文件生成對(duì)應(yīng)的口型和表情。

這里需要注意的是，音頻文件只支持MP3或者wav。

除此以外，推理過(guò)程中Pytorch庫(kù)可能會(huì)報(bào)這個(gè)錯(cuò)誤：

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 6.00 GiB total capacity; 5.38 GiB already allocated; 0 bytes free; 5.38 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

這就是所謂的"爆顯存問(wèn)題"。

一般情況下，是因?yàn)楫?dāng)前GPU的顯存不夠了所導(dǎo)致的，可以考慮縮小torch分片文件的體積：

set PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:60

如果音頻文件實(shí)在過(guò)大，也可以通過(guò)ffmpeg對(duì)音頻文件切片操作，分多次進(jìn)行推理：

ffmpeg -ss 00:00:00 -i test_en.wav -to 00:30:00 -c copy test_en_01.wav

藉此，就解決了推理過(guò)程中的爆顯存問(wèn)題。

和Wav2Lip相比，SadTalker(Stylized Audio-Driven Talking-head)提供了更加細(xì)微的面部運(yùn)動(dòng)細(xì)節(jié)（如眼睛眨動(dòng)）等等，可謂是細(xì)致入微，巨細(xì)靡遺，當(dāng)然隨之而來(lái)的是模型數(shù)量和推理成本以及推理時(shí)間的增加，但顯然，這些都是值得的。

標(biāo)簽：