FreeSWITCH 简单图形化界面43 - 使用百度的unimrcp搞个智能话务台，用的在线的ASR和TTS

---

0、一个fs的web配置界面预览

http://myfs.f3322.net:8020/
用户名：admin，密码：admin

FreeSWITCH界面安装参考：https://blog.csdn.net/jia198810/article/details/137820796

1、安装unimrcp模块

参考链接：
1、https://docs.qq.com/pdf/DVFFub1JpSmNMRXdJ] 百度mrcp章节
2、https://blog.csdn.net/jia198810/article/details/118461176?spm=1011.2415.3001.5331

安装完成后，配置unimrcp服务器的端口为6060。

如果这里模块没有安装成功，就不用继续了哈，请划过。

2、安装完成后，配置FreeSWITCH。

2.1 有界面的配置

2.1.1 mod_unimrcp模块配置

2.1.2 mod_unimrcp客户端配置

2.2 无界面的配置

如果使用的是没有界面的FreeSWITCH，配置如下：

<configuration name="unimrcp.conf" description="UniMRCP Client"><settings><!--  UniMRCP profile to use for TTS  --><param name="default-tts-profile" value="baidu"/><!--  UniMRCP profile to use for ASR  --><param name="default-asr-profile" value="baidu"/><!--  UniMRCP logging level to appear in freeswitch.log.  Options are:EMERGENCY|ALERT|CRITICAL|ERROR|WARNING|NOTICE|INFO|DEBUG  --><param name="log-level" value="info"/><!--  Enable events for profile creation, open, and close  --><param name="enable-profile-events" value="true"/><param name="max-connection-count" value="100"/><param name="offer-new-connection" value="1"/><param name="request-timeout" value="3000"/></settings>
<profiles>
<profile name="baidu" version="2"><param name="server-ip" value="127.0.0.1"/><!--  Mrcp Server 端口号  --><param name="server-port" value="6060"/><param name="resource-location" value=""/><param name="client-ip" value="127.0.0.1"/><param name="client-port" value="6061"/><param name="sip-transport" value="udp"/><param name="rtp-ip" value="127.0.0.1"/><param name="rtp-port-min" value="30000"/><param name="rtp-port-max" value="30500"/><param name="speechsynth" value="speechsynthesizer"/><param name="speechrecog" value="speechrecognizer"/><param name="codecs" value="PCMU PCMA L16/96/8000"/><!--  Add any default MRCP params for SPEAK requests here  --><synthparams> </synthparams><!--  Add any default MRCP params for RECOGNIZE requests here  --><recogparams><!-- param name="start-input-timers" value="false"/ --></recogparams></profile>
</profiles>
</configuration>

3、呼叫规则

<extension name="测试:11120"><condition expression="^(11120)$" field="destination_number"><action application="python" data="baidu-mrcpserver"/><action application="hangup"/></condition>
</extension>

如下图：

4、编写流程

编写baidu-mrcpserver.py代码，放在/usr/local/freeswitch/scripts/下面，如下代码用的python，lua脚本换下语法即可，Freeswitch的lua API和python的API是通用的，名称都一样。

4.1 python代码

#!/usr/local/python3/bin/python3
from freeswitch import *
"""
baidu-mrcpserver.py
拨号规则:
<extension name="测试:11120"><condition expression="^(11120)$" field="destination_number"><action application="python" data="baidu-mrcpserver"/><action application="hangup"/></condition>
</extension># 识别用户输入后，交由第三方处理，并读出结果。
"""
from xml.dom import minidom
from openai import OpenAI
import redeepseek_api_key = "i am a key "# 解析百度的asr结果,XML格式的
def parse_xml(xml_string):text = ""try:dom = minidom.parseString(xml_string)input_element = dom.getElementsByTagName('input')[0]text = input_element.firstChild.nodeValueexcept Exception as e:consoleLog("error", f"解析xml出错:{e}")return text# 说
def speak(session, text):if text:session.consoleLog("info", f"正在说：{text}")session.execute("speak", f"unimrcp:baidu||{text}")# 开始听
def start_listen(session):# timeout为语音识别时间,单位为毫秒# session.consoleLog("info", f"开始语音识别")# 设置识别超时时间timeout = 10000session.execute("detect_speech", f"param speech-timeout {timeout}")session.execute("detect_speech", "unimrcp baidu builtin:grammar/baidu")# 暂停听
def pause_listen(session):session.consoleLog("info", "暂停语音识别")session.execute("detect_speech", "pause")# 结束听
def stop_listen(session):session.consoleLog("info", "结束语音识别")session.execute("detect_speech", "stop")# 恢复听
def resume_listen(session):session.consoleLog("info", "恢复语音识别")# resume 不好使,用detect_speech代替# session:execute("detect_speech", "resume");start_listen(session)# 连说带听,不会触发回调
def speak_and_listen(session, text, time=5000):session.execute("set","tts_engine=unimrcp")session.execute("play_and_detect_speech",f"say:{text} detect:unimrcp:baidu builtin:grammar/baidu")session.execute("sleep", f"{time}")# 第三方处理
# 这里可以编写对于语音识别的处理,这里是调用deepseek的python接口处理语音识别结果
def process_result(session, result):if "挂机" in result or "退出" in result or "结束" in result or "再见" in result or "结束通话" in result or "挂断" in result or "拜拜" in result:speak(session, "再见")session.hangup()return "break"if session.ready():speak(session, f"正在查询{result},请稍等")try:sentence_delimiters = r'([，。！？,\.!\?\n])'  # [1,3,7](@ref)buffer = ""current_sentence = ""client = OpenAI(api_key=deepseek_api_key, base_url="https://api.deepseek.com")response = client.chat.completions.create(model="deepseek-chat",messages=[{"role": "system", "content": "生活助手"},{"role": "user", "content": f"{result},不要超过100字回复,简洁回答,不要出现表情、星号、井号等特殊符号"},],stream=True  # 设置 stream 参数为 True)for chunk in response:if chunk.choices and chunk.choices[0].delta.content:content = chunk.choices[0].delta.contentbuffer += content# 实时检测分隔符while True:match = re.search(sentence_delimiters, buffer)if not match:break# 分割出完整句子split_pos = match.end()sentence_part = buffer[:split_pos]buffer = buffer[split_pos:]# 将句子片段添加到当前句子current_sentence += sentence_part# 判断是否遇到句子结束符if match.group() in ['。', '.', '！', '!', '？', '?', '\n']:  # [3,7](@ref)# 分段说,避免句子太长,tts生成的太慢speak(session,current_sentence.strip())current_sentence = ""# 输出缓冲区剩余内容if buffer.strip():speak(session,current_sentence.strip())except Exception as e:consoleLog("info", f"处理结果出错:{e}")return "break"# 只有在detect_speech成功后才会调用,play_and_detect_speech不会调用
# 最大识别次数,超过次数挂机
max_retries = 3
def on_input(session, callback_type, obj, arg):global max_retriesif max_retries <= 0:# 重试次数用完,结束监听speak(session, "通话超时，感谢您的使用")session.hangup()return "break"# 处理用户按键if callback_type == "dtmf":session.consoleLog("info", f"用户按了DTMF,收集到的按键:{obj.digit}")return ""# 处理事件if callback_type == "event":event_type = obj.getHeader("Speech-Type")# 触发了detected-speech事件if event_type == "detected-speech":speech_output = obj.getBody()if speech_output:text = parse_xml(speech_output)if text and len(text) >= 2:# 获取结果后进行其他处理process_result(session, text)speak(session, "还有什么可以帮到您?")else:# 识别结果太短,重新开始监听speak(session, "没听清,请再说一遍")else:# 识别结果太短,重新开始监听speak(session, "没听清,请再说一遍")max_retries -= 1# 立即结束本次处理return "break"# 立即结束本次处理return "break"# 立即结束本次处理return "break"# 通话入口
def handler(session, args):# 应答session.answer()# 是否向event_socket发送asr事件,我们是拨号规则处理asr事件,所以不需要# session.setVariable("fire_asr_events", "true")# 注册事件回调,python的回调函数不要加引号session.setInputCallback(on_input)# 欢迎语speak(session,text="请问有什么可以帮助您的?")# 一直听,识别,读结果,直到max_retries用完while session.ready():start_listen(session)session.streamFile("silence_stream://90000000")# 挂机session.hangup()

4.2 主要是detect_speech和speak

5、效果

freeswitch 百度mrcp测试

6、 一些参考的官网网址

mod_unimrcp
https://developer.signalwire.com/freeswitch/FreeSWITCH-Explained/Modules/mod_unimrcp_6586728
MRCP的参数设置
No-Input-Timeout, Recognition-Timeout, Start-Input-Timers
https://www.rfc-editor.org/rfc/rfc6787
speak
https://developer.signalwire.com/freeswitch/FreeSWITCH-Explained/Modules/mod-dptools/6587123/#0-about
play_and_detect_speech
https://developer.signalwire.com/freeswitch/FreeSWITCH-Explained/Modules/mod-dptools/6586714#lua
detect_speech
https://developer.signalwire.com/freeswitch/FreeSWITCH-Explained/Modules/mod-dptools/6586756#about
详细的语音识别流程,识别用户分机的案例
https://developer.signalwire.com/freeswitch/FreeSWITCH-Explained/Client-and-Developer-Interfaces/Lua-API-Reference/Lua-examples/Lua-ASR-TTS-Directory-example_1049011/

祝君成功，好运连连