CN111901488B

CN111901488B - Method for improving outbound efficiency of voice robot based on number state

Info

Publication number: CN111901488B
Application number: CN202010672764.0A
Authority: CN
Inventors: 李训林; 王帅
Original assignee: Shengzhi Information Technology Nanjing Co ltd
Current assignee: Shengzhi Information Technology Nanjing Co ltd
Priority date: 2020-07-14
Filing date: 2020-07-14
Publication date: 2021-11-09
Anticipated expiration: 2040-07-14
Also published as: CN111901488A

Abstract

The invention discloses a method for improving the outbound efficiency of a voice robot based on number states, which relates to the technical field of intelligent voice interaction; after the call bridging is successful, the invention releases the resources of the invalid line by identifying the call state, thereby improving the AI outbound efficiency; the identification and processing of the state of the outbound number can be completed in real time by dynamic configuration according to the characteristics of the regional scene without excessively collecting linguistic data or depending on operators. And the voice mailbox and the balance playing state after answering can be effectively solved.

Description

Method for improving outbound efficiency of voice robot based on number state

Technical Field

The invention relates to the technical field of intelligent voice interaction, in particular to a method for improving the outbound efficiency of a voice robot based on number states.

Background

The traditional electricity sales industry is a highly human-input industry, and electricity sales servicers are faced with the same invitations and frequently complaints and abuses of customers in the face of the same invitations and different people in a large number of daily outcalls. Aiming at the situations, the working state of the customer service is easily influenced by subjective emotion, and an uncontrollable state is achieved. With the development of artificial intelligence technology, intelligent outbound robots based on a voice system come out, and great opportunity is brought to the electricity selling industry. The system can simulate real person communication, easily deal with multi-turn dialogue response, actively guide the sales process and achieve the purpose of screening the intention customers.

The intelligent voice robot guides the client to carry out conversation in a Torontal conversation mode on the premise of presetting a conversation scene, so that the marketing purpose is achieved. Its main core function modules Are Speech Recognition (ASR), speech synthesis (TTS), Dialogue Management (DM), Natural Language Processing (NLP), and external calling system (CHS).

The AI outbound system is a core component module of the intelligent voice robot, and the main core responsibility of the AI outbound system is to outbound the number according to the requirements of the client; the specific process is as follows: 1. the client system module initiates a batch outbound request to the AI outbound system; 2. after receiving the request of the client system, the outbound system module calls the gateway system to initiate batch calls to the clients according to the maintained number of the outbound lines; 3. after the other party answers, the dialogue management system carries out dialogue with the client through the outbound system; 4. after the call is completed, the called party hangs up the call to finish the call or the conversation management system triggers an on-hook instruction to inform an outbound system to hang up to finish the call; 5. after the call is over, the outbound system will continue to call the next customer number in the outbound queue provided by the customer system.

However, in the real outbound process of the intelligent voice robot, a large number of invalid telephones (the invalid numbers mainly refer to telephones which cannot reach customers due to the fact that a calling party is empty, the calling party is stopped, the called party is stopped, the intelligent voice robot is powered off, is in a conversation, is not in a service area, is busy, plays balance, voice mail and the like) exist, and the telephones which are invalid according to statistics occupy 60% of the total daily outbound call volume of the outbound system. If the invalid telephone is automatically hung up, the intelligent voice robot can be hung up only when the ringing is finished according to the analysis. Suppose that the time of each ringing tone is 60s, that is, each telephone AI needs to wait for the operator to hang up, that is, the AI can stop calling the number after 60s, and then releases the line to call the next number. Therefore, when a large number of outbound numbers of the type appear, the outbound efficiency of the AI is seriously influenced, and the resource is seriously wasted. During the whole call process, the calling line of the outbound system is always occupied by resources. If the called party is in other states such as shutdown, no one answers or the called party stops, and the outbound system cannot acquire the on-hook instruction, the line resource can be released only by waiting for the operator to stop the call, and the next customer number call is started, which results in the waste of the line resource.

Aiming at the current problems, the current industry has the following solutions:

1. and the outbound system inquires the home operator of the number according to the outbound number. After the call is finished, the ring-back tone and the check code fed back by the operator are searched and compared in the operator number state base corresponding to the preset operator number. Since this state needs to be acquired after the call is ended, there is no help to the AI outbound efficiency.

2. Collecting the voice of each operator area, making an audio frequency library, converting the analog voice signal after the call-out into digital information through a voiceprint recognition technology or a sample library, and comparing the digital information with the sample library to realize the recognition of the number state. The method has the problems that a large amount of linguistic data needs to be collected, a large amount of research and development cost exists, and the input-output ratio is not high. On the other hand, the comparison requires processing a large amount of data, which wastes resources and consumes time. On the other hand, when the product enters a new market, such as nation a, such corpus collection is a difficult and important factor for us. If the voice enters B, a large amount of voice samples need to be collected continuously. Therefore, high reusability of functions becomes an important factor;

at present, the general practice in the whole intelligent voice industry is to convert the audio from the voice signal to the digital signal when the called party is not listening and is in the broadcasting voice state of the operator, and then compare the digital signal with the audio corpus collected in advance to judge the number state. But the whole scheme has lower input and output; firstly, in the data collection stage, the prompt tone is influenced by the difference between operators and regions, and different operators or regions may have different prompt tones; the collection of the prompt audio needs a lot of manpower and is difficult to collect completely. Secondly, in the data use stage, when the corpus of the audio library is too much, the comparison process needs a large amount of operations, and the overall efficiency is low. When the product covers the area without audio collection, the technical support is difficult to keep up with the business expansion; the overall function is not flexible enough; therefore, how to correctly and efficiently identify each invalid telephone is a key problem to be solved at present, and the method can be used across languages and regions, so that the research on the invalid telephone identification mode has practical significance.

Disclosure of Invention

The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method for improving the outbound efficiency of a voice robot based on a number state, and the outbound efficiency of the voice robot is improved by adopting a flexible and extensible way to identify the number state.

The invention adopts the following technical scheme for solving the technical problems:

the method for improving the outbound efficiency of the voice robot based on the number state comprises the following steps:

step 1, configuring an invalid number recognition rule through an intelligent voice outbound management system;

step 2, selecting an invalid number identification rule to be used by taking a call task as a dimension, wherein all calls under the task are used as identification logic for invalid number verification based on the invalid number identification rule;

step 3, after the client system module initiates a request call to the outbound system, the outbound system calls a gateway to initiate a call to the called number; after the call is established, the intelligent voice outbound management system detects voice through a voice endpoint detection VAD function and continuously sends a reception result to the feature extraction module;

step 4, the feature extraction module sends the language and the received audio to a voice recognition module for voice processing according to the language information maintained by the task;

step 5, after the speech recognition module receives the language and the audio, performing speech-to-text recognition on the received speech; after the recognition is finished, the voice recognition module returns the final text to the feature extraction module;

step 6, the feature extraction module acquires a final text returned by the voice recognition module, judges the current calling state, and performs keyword and regular matching on the final text and classification characteristics of invalid numbers before answering configured in the intelligent voice outbound management system if the current calling state is before user response; if the configured classification characteristics are matched, the number is considered as an invalid number; the feature extraction module initiates a notice to the outbound system, requests to terminate the call, and reports invalid numbers to the outbound system; after the outbound system receives the instruction, the calling state is recorded as an invalid number, and the call is actively hung up; starting to call the next queue number;

step 7, if the classification is not matched, the feature extraction module continues to monitor the broadcast, and carries out clipping identification and feature extraction operation until the operator responds successfully; the outbound system informs the dialogue management module that the called party answers; the dialogue management module informs the outbound system to broadcast corresponding dialogue;

step 8, after response, in the speech broadcasting process, the outbound system module continues to monitor the user speaking, if the voice is collected, the voice is edited and sent to the voice recognition module when the voice is silent;

step 9, the voice recognition module carries out voice-to-text recognition on the received voice; after the recognition is finished, the voice recognition module returns the final text to the feature extraction module;

step 10, a feature extraction module acquires a text returned by a voice recognition module, and keywords and regular matching are carried out on the text and text characteristics corresponding to invalid number classification after user response configured in an intelligent voice outbound management system; according to the matching rule, the feature extraction module judges whether the telephone is an invalid number in a voice mailbox or a balance playing scene; if the voice mailbox or the balance is played, the feature extraction module informs the outbound system that the number is the invalid number in the scene of voice information or balance playing, and the outbound system sends the state of the text and the invalid number recognized by the voice recognition system to the dialogue management module; after the dialogue management module acquires the text and number states of the outbound system, the judgment is carried out: 1. if the voice mail box is the voice mail box, returning the voice mail box corresponding to the talk, and informing the outbound system to execute hang-up operation after the talk playing of the voice mail box is finished; 2. if the balance is played, the hang-up request of the outbound system is directly sent.

As a further optimization scheme of the method for improving the outbound efficiency of the voice robot based on the number state, the invalid number recognition rule in step 1 is as follows: firstly, combing language, region and response state indexes, collecting responses of invalid number classification corresponding to index information, and extracting characteristic keywords or regular characters according to the invalid number classification; and then obtaining response of invalid number classification according to language, region and response state.

Compared with the prior art, the invention adopting the technical scheme has the following technical effects:

(1) after the call bridging is successful, the invention releases the resources of the invalid line by identifying the call state, thereby improving the AI outbound efficiency;

(2) the identification and processing of the state of the outbound number can be completed in real time by dynamic configuration according to the characteristics of the regional scene without excessively collecting linguistic data or depending on operators. And the voice mailbox and the balance playing state after answering can be effectively solved.

Drawings

FIG. 1 is a schematic flow diagram of the present invention.

FIG. 2 is a feature extraction module.

Detailed Description

The technical scheme of the invention is further explained in detail by combining the attached drawings:

fig. 1 is a schematic flow diagram of the present invention, specifically as follows:

the method comprises the following steps: preparation work

The configuration state of the feature extraction module of the intelligent voice robot is classified, and the configuration state of the feature extraction module can be classified according to the environment of an operator where the market is located. For example, the classification of singapore configuration is as follows: busy, Reject, Caller down, Called down, Empty, Shutdown;

and marking the classified state as before answering or after answering;

step two: preparation work

According to the classification state, collecting the classification sentence text, and marking the classification text characteristics, as shown in the invalid number text and state corresponding relation table 1,

TABLE 1

Step three: preparation work

Configuring the type of the enterprise voice engine according to the language of the enterprise market environment, and configuring an English engine ASR for the voice engine of the enterprise if the affiliated example market is an English environment;

step four:

the method comprises the steps that a batch of outbound client numbers are led in from a client management system, the client management system initiates call requests to an outbound system in sequence according to the numbers needing outbound and outbound concurrency numbers of a task (if the numbers are ABC three client numbers, and the number of the outbound concurrency numbers is 1, the client system sends the A numbers to the outbound system);

step five:

after the calling system receives the calling request of the client system, the calling system calls the calling number of the gateway to call the client calling number. Namely, a call is initiated to the A number;

step six:

the outbound system carries out endpoint detection on the broadcast of early media through a built-in VAD module, and a mute threshold value is configured in advance to be 200 ms.

When the calling system requests the gateway to successfully call, the VAD detects the broadcast or ring back tone of the operator, and when the broadcast or ring back tone of the operator has a mute interval of 200ms, the broadcast segment of the previous sentence is cut, including muting for 200 ms. And send the voice slice to the voice recognition module.

Considering that the operator's null number audio playback is generally interrupted, the null number audio is generally not available at one time, and therefore the outbound call system needs to continue to perform subsequent radio reception and slicing on the broadcast through the VAD module.

After the speech recognition module returns the text result of ASR recognition, the result is given to the feature extraction module, and the state of the call is marked as before answering;

step seven:

the feature extraction module is shown in fig. 2, and the result text X according to ASR is as follows: the number you have a able of being able to answer at The moment, and The state before answering, and The character of The classified text of The state before answering configured in advance are matched by means of key words, regular patterns and The like, if The corresponding text sentence is matched, The classification of The sentence such as Shutdown is output, and The system is informed to The outbound system, at this moment, The outbound system actively hangs up The call, and The call result is returned to The client system. The client system will continue to call the B client request to the outbound system;

and if the text returned by the ASR is Y and is matched with the pre-configured classification text characteristics, the corresponding text classification is not matched, and the outbound system continues slicing to continue the process. Until the called answering operator returns 200ok or the characteristic state is output for classification;

step eight:

and when the operator returns 200ok in the normal scene, the operator answers the call on behalf of the user, and at the moment, the outbound call system informs the conversation management module that the called party answers the call. The called management module informs the outbound system of a broadcasting node, namely Script M, and after the broadcasting of the Script M is finished, the outbound system collects the voice of the client, detects the voice in the same way, receives the voice if the voice exists and clips the voice when the voice is silent. Sending the sliced voice to ASR of a corresponding engine for voice recognition;

if the voice is not collected, the dialogue management module is informed, the user responds to the overtime, and the dialogue management module informs the outbound system of broadcasting the preset Script N according to the overtime. And then the user responses do not carry out text classification feature extraction.

Due to the particularity of the voice mailbox, the voice mailbox is generally an operator which automatically answers the call and then automatically broadcasts the voice mailbox. So when the operator returns 200ok, which does not represent the true user answering, it may be a voice mailbox;

therefore, in response to the user response of the first node, the ASR returns the text, and then sends the text to the feature extraction module. And matching the characters of the preset classified texts in a way of keywords, regularization and the like, and if the corresponding text sentences are matched, outputting the corresponding classification to an outbound system. The outbound system makes an active on-hook. If no classification is identified, the system determines that the mailbox is not voice, and sends the text to the dialog management module, which further determines the broadcast telephony Script XX.

The above description is only for the specific embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.

Claims

1. A method for improving the outbound efficiency of a voice robot based on number state is characterized by comprising the following steps:

2. The method for improving the outbound efficiency of the voice robot based on the number status according to claim 1, wherein the invalid number recognition rule in step 1 is as follows: firstly, combing language, region and response state indexes, collecting responses of invalid number classification corresponding to index information, and extracting characteristic keywords or regular characters according to the invalid number classification; and then obtaining response of invalid number classification according to language, region and response state.