基于LLM模型实现与metahuman克隆数字人对话，过程总结

[wxzz]

前面文章参考：克隆人像和声音，metahuman与GPT-SoVITS接口打通，过程汇总（https://github.com/lipku/metahuman-stream/issues/210）

1.流程概要：前台语音及上传到后台->asr语音识别->LLM对话模型->metahuman呈现 2.增加asr语音识别库 下载模型：https://openaipublic.azureedge.net/main/whisper/models/9ecf779972d90ba49c06d968637d720dd632c55bbf19d441fb42bf17a411e794/small.pt 把模型名称改为：whisper_small.pt，在metahuman工程根目录新建目录“sound”,并把模型文件复制到“sound”目录下。

在“sound”目录下新建：speech2text.py 程序文件，代码实现如下：
import os
import whisper 

os.environ["HF_ENDPOINT"] = "https://hf-mirror.com" os.environ["CUDA_VISIBLE_DEVICES"] = "2" os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"

model = whisper.load_model("sound/whisper_small.pt") #['tiny.en', 'tiny', 'base.en', 'base', 'small.en', 'small', 'medium.en', 'medium', 'large-v1', 'large-v2', 'large-v3', 'large']

def speech2text(speech_file): result = model.transcribe(speech_file,fp16=False,verbose = True,language="Chinese") return result['text']

3.增加LLM对话模型，我们使用百度的文心一言 在llm目录增加：yiyan_sdk.py 程序文件，代码实现如下： import os import qianfan

os.environ["QIANFAN_AK"] = "fLdNo" os.environ["QIANFAN_SK"] = "mXjXq"

def yiyan_chat(user,msg): resp = qianfan.ChatCompletion().do(endpoint="completions_pro", messages=[{"role":""+user+"","content":""+msg+""}], temperature=0.95, top_p=0.8, penalty_score=1, disable_search=False, enable_citation=False) return resp.body["result"]

4.修改前端界面：webrtcapi.html，增加代码。

录制对话

<button id="upload_sound_btn">对话</button>
<audio class="mt-1" style="width: 300px;height: 40px;" id="audioPlayer" controls></audio>

日志：

5.修改js操作代码：client.js，增加代码如下： let mediaRecorder; let audioChunks = []; const chatButton = document.getElementById('chat_btn'); const audioPlayer = document.getElementById('audioPlayer'); chatButton.addEventListener('mousedown', start_sound); chatButton.addEventListener('mouseup', stop_sound);

const uploadRecordButton = document.getElementById('upload_sound_btn');
uploadRecordButton.addEventListener('click', upload_sound);

const chatlogText = document.getElementById('chatlog');

function start_sound() { audioChunks = []; navigator.mediaDevices.getUserMedia({audio: true}) .then(stream => { mediaRecorder = new MediaRecorder(stream); mediaRecorder.ondataavailable = event => { if (event.data.size > 0) { audioChunks.push(event.data); } }; mediaRecorder.onstop = () => { const audioBlob = new Blob(audioChunks, {type: 'audio/wav'}); const audioUrl = URL.createObjectURL(audioBlob); audioPlayer.src = audioUrl; }; mediaRecorder.start(); chatButton.textContent = '录制中...'; audioPlayer.src = null; audioPlayer.load(); //startRecordButton.disabled = true; //stopRecordButton.disabled = false; }) .catch(error => { console.error('getUserMedia error:', error); }); }

function stop_sound() { mediaRecorder.stop(); //startRecordButton.disabled = false; chatButton.textContent = '录制对话'; //stopRecordButton.disabled = true; }

function upload_sound() { //上传对话音频 if(audioPlayer.src==null) return ; const audioBlob = new Blob(audioChunks, {type: 'audio/wav'}); const formData = new FormData(); formData.append('audio', audioBlob,'recorded_audio.wav'); sessionid=document.getElementById('sessionid').value; formData.append('sessionid',parseInt(sessionid)); let index = layer.load(); fetch('/uploadsound', { method: 'POST', body: formData }) .then(response => response.json()) .then(data => { layer.close(index);

        chatlogText.value += data['data'] + '\n';
    })
    .catch(error => {
        layer.close(index);
        console.error('Upload error:', error);
});

}

6.修改后端代码 在if name == 'main'函数中增加： appasync.router.add_route(method='POST',path='/uploadsound',handler=uploadsound)

增加uploadsound函数，代码实现如下： async def uploadsound(request):

try:
    data= await request.post()
    sessionid = int(data['sessionid'])
    fileobj = data["audio"]
    filename=fileobj.filename
    dir=os.path.join("sound")
    if(not os.path.exists(dir)):
        os.mkdir(dir)

    save_dir=os.path.join(dir,filename)

    if(os.path.exists(save_dir)):
        os.remove(save_dir)

    filebytes=fileobj.file.read()

    with open(save_dir,'wb') as file:
        file.write(filebytes)

    text=speech2text(save_dir)

    text_respose=yiyan_chat("user",text)

    nerfreals[sessionid].pause_talk()
    nerfreals[sessionid].put_msg_txt(text_respose)

    return web.Response(
        content_type="application/json",
        text=json.dumps(
            {"code": 0, "msg":"ok","data": ""+"【我】:"+text+"\n"+"【数字我】:"+text_respose+""}
        ),
    )
except Exception as e:
    return web.Response(
        content_type="application/json",
        text=json.dumps(
            {"code": -1, "msg":"err","data": ""+e.args[0]+""}
        ),
    )

7.实现效果，如下图

[JerryDaHeLian]

大赞，可以分享下github代码吗？

[xiaomaofei]

大佬有错误，有报错。分享一下

[wxzz]

大赞，可以分享下github代码吗？

完整代码就在上面。

[wxzz]

大佬有错误，有报错。分享一下

有错误，分享啥。