ASR(Automatic Speech Recognition)自动语音识别技术是一种将人的语音转换为文本的技术。
一、freeswitch如何使用asr
freeswitch提供两个app功能detect_speech和play_and_detect_speech给用户调用,detect_speech是异步的,play_and_detect_speech是同步的。
1、detect_speech
1.1语法:detect_speech <mod_name> <gram_name> <name> <dest>
detect_speech stop
mod_name: 识别模块名称 如ali_asr[:params],其中params是param1=value1,param2=value2类型格式传入到switch_asr_interface_t接口中asr_open接口函数
gram_name:switch_asr_interface_t接口中asr_load_grammar接口函数传入参数
name: switch_asr_interface_t接口中asr_load_grammar接口函数传入参数
dest:switch_asr_interface_t接口中asr_open接口函数传入参数
stop:停止识别
1.2识别事件:设置通道变量fire_asr_events为true。freeswitch会触发SWITCH_EVENT_DETECTED_SPEECH事件返回给事件回调函数。
1.2.1事件头Speech-Type:
begin-speaking:检测到用户说话
detected-speech:识别结果
closed:识别关闭
1.2.2 事件body:识别结果
备注:事件可以用switch_event_bind("mod_asr", SWITCH_EVENT_DETECTED_SPEECH, NULL, my_event_handler, NULL)绑定;
2、play_and_detect_speech
2.1语法 play_and_detect_speech <file> detect:<engine>[:<mrcp_profile>] {param1=val1,param2=val2}<grammar>
file:语音文件或者tts
engine:识别模块
mrcp_profile:mrcp文件
{param1=val1,param2=val2}<grammar> 参数+语法文件
2.2用户可以通过app返回通道变量detect_speech_result获取识别结果
综上所述,异步函数detect_speech可以通过事件SWITCH_EVENT_DETECTED_SPEECH获取识别结果,同步函数 play_and_detect_speech既可以交互语音还能获取识别结果。
二、如何实现一个自定义的asr模块
众所周知:市面上的asr对接的中文识别引擎有阿里,科大讯飞,百度,想在fs中实现对接有3种方式:
1、mrcp对接,fs提供mod_unimrcp模块通过mrcp对接各类厂家asr服务,前提是厂家支持这种协议。具体细节可以参考官网的例子mod_unimrcp - FreeSWITCH - Confluence
2、通过实现media bug接口来实现
我们可以通过switch_core_media_bug_add(session, "my_detect_speech", key, speech_callback, sth, 0, SMBF_READ_STREAM | SMBF_NO_PAUSE | SMBF_ONE_ONLY, &sth->bug))来绑定识别回调函数,我们只要简单的实现回调函数来获取语音流媒体,把流媒体发送给识别引擎就获取识别结果。
static switch_bool_t speech_callback(switch_media_bug_t *bug, void *user_data, switch_abc_type_t type)
{
fastcall_speech_handle_t *sth = (fastcall_speech_handle_t *) user_data;
uint8_t data[SWITCH_RECOMMENDED_BUFFER_SIZE];
switch_frame_t frame = { 0 };
switch_asr_flag_t flags = SWITCH_ASR_FLAG_DATA;
frame.data = data;
frame.buflen = SWITCH_RECOMMENDED_BUFFER_SIZE;
switch (type) {
case SWITCH_ABC_TYPE_INIT:
{
}
break;
case SWITCH_ABC_TYPE_CLOSE:
{
}
break;
case SWITCH_ABC_TYPE_READ:
if (sth->ah) {
if (switch_core_media_bug_read(bug, &frame, SWITCH_FALSE) != SWITCH_STATUS_FALSE) {
//frame.data, frame.datalen是我们需要的语音数据
}
}
}
break;
case SWITCH_ABC_TYPE_WRITE:
default:
break;
}
return SWITCH_TRUE;
}
3、通过fs现有ASR接口方式
asr_interface = (switch_asr_interface_t *)switch_loadable_module_create_interface(*module_interface, SWITCH_ASR_INTERFACE);
asr_interface->interface_name = "fastcall_asr";
asr_interface->asr_open = fastcall_asr_open;
asr_interface->asr_load_grammar = fastcall_asr_load_grammar;
asr_interface->asr_unload_grammar = fastcall_asr_unload_grammar;
asr_interface->asr_close = fastcall_asr_close;
asr_interface->asr_feed = fastcall_asr_feed;
//asr_interface->asr_feed_dtmf = fastcall_asr_feed_dtmf;
asr_interface->asr_resume = fastcall_asr_resume;
asr_interface->asr_pause = fastcall_asr_pause;
asr_interface->asr_start_input_timers = fastcall_asr_start_input_timers;
asr_interface->asr_check_results = fastcall_asr_check_results;
asr_interface->asr_get_results = fastcall_asr_get_results;
asr_interface->asr_text_param = fastcall_asr_text_param;
asr_interface->asr_numeric_param = fastcall_asr_numeric_param;
asr_interface->asr_float_param = fastcall_asr_float_param;
我们只需要实现对应的函数就可以获取到识别结果。
三、软件架构设计
1、编写mod_AI模块实现Freeswitch的ASR和TTS接口。
2、编写具有插件功能的ASR_Server系统,生成接口
struct fsl_asr_interface {
/*! the name of the interface */
const char *interface_name;
/*! function to open the asr interface */
int (*asr_open) (fsl_asr_handle_t *ah, char *param, int realtime);
/*! function to close the asr interface */
int (*asr_close) (fsl_asr_handle_t *ah);
int (*asr_start_session) (fsl_asr_handle_t* ah, int rate, char *param);
int (*asr_stop_session) (fsl_asr_handle_t* ah);
/*! function to feed audio to the ASR */
int (*asr_feed) (fsl_asr_handle_t *ah, void *data, unsigned int len, int audio_state, int* ep_stat, int* rec_stat);
int (*asr_realtime_feed) (fsl_asr_handle_t* ah, void* data, unsigned int len, char** rst);
/*! function to read results from the ASR */
int (*asr_get_result)(fsl_asr_handle_t *ah, char **rst);
fsl_thread_rwlock_t *rwlock;
int refs;
fsl_mutex_t *reflock;
fsl_loadable_module_interface_t *parent;
struct fsl_asr_interface *next;
};
给插件使用,方便扩展。
3、分别编写阿里,讯飞,百度模块提供识别引擎功能
4、ASR_Server通过mod_AI传送过来的参数interface_name来调用对应的模块进行识别。