Voiceprint recognition helps you keep your money bag, talk about artificial intelligence in the securities industry

Financial transactions such as stocks and securities are increasingly moving away from offline to online. Biometric technologies such as fingerprints and human faces are popularized. How to accurately identify and authenticate "money"-related scenes? How to ensure business compliance? in particular:

• When trading online, how to ensure that the trader who is buying and selling is himself?
• Remote control, how to ensure the safety of funds for traders?
• How can the voiceprint recognition technology, which has the advantages of "non-contact" and "remote recognition", be implemented in business scenarios?

Last Saturday, the Arch Meetup Shenzhen station hosted by the Milvus community came to a successful conclusion. Many technology enthusiasts gathered together. Lecturers from Zilliz, Zhuiyi Technology, Palm Data Technology, GitLab and ShowMeBug shared with you The latest trends and application scenarios of open source software. Palm Data Technology Director Gao Xing shared the needs and pain points of the securities and fund industry, as well as "question and answer robots" and "voiceprint recognition" and other financial and securities industry operational tools. How does the open source vector database Milvus help realize the above scenarios? Let's see it together!

To tell you quietly, follow the Zilliz public account and reply "ppt9" get wonderful sharing from the other four lecturers!

Palm Data Technology x Milvus Vector Database, what sparks will it spark in the financial AI field?

Hi-tech, technical director of Palm Data Technology, shared the scene

Palm Data Technology focuses on the securities and large asset management industries, and is a technology company that manages data security in the financial industry for the fields of big data and AI basic technology platforms, intelligent scene applications and data security management. Palm Data Technology and Xiamen University established the "Financial Technology Joint Laboratory", dedicated to the application research of big data and artificial intelligence technology in the field of financial technology. This year, Palm Data Technology also officially became one of the five application pilot units of the national standard "Information Security Technology Voiceprint Identification Data Security Requirements" of the WG4 Voiceprint Group.

In the era of Embedding, pictures, text, video, voice and other unstructured data can be extracted as feature vectors through Embedding technology, and then through the calculation of feature vectors and similarity retrieval to help implement intelligent question answering, product recommendation, voice Pattern recognition and other scenes. The open source vector database Milvus can empower AI applications and vector similarity searches. The open-source vector database Milvus supports the use of multiple AI models to vectorize unstructured data and provides search and analysis services for vector data. The services it can handle include image processing, machine vision, natural language processing, speech recognition, recommendation systems, and new drug discovery. The specific implementation is:

• Convert unstructured data into feature vectors through deep learning models and import them into Milvus database;
• Store and index the feature vector;
• After receiving the vector search request from the user, it returns a result similar to the input vector.

Based on years of industry experience, Palm Data Technology uses the Milvus vector database to achieve rapid response in scenarios such as "Question and Answer Robot" and " " through 1617bdb46d7fae "Database + Application" , Easy to expand, low cost, easy to maintain solutions.

The needs and trends of the financial securities industry

Finance is a wealth-concentrated field. In the securities market that is the main focus of Palm Data Technology, its total assets will reach 8.9 trillion in 2020. It has the characteristics of a high proportion of retail investors, large fluctuations due to many factors, and a small number of financial products. Therefore, the customer operation field of the securities industry naturally produced the following two requirements:

(1) How to serve investor customers efficiently, reliably and in compliance with the rapid growth of user scale?

• The financial securities industry needs to improve the level of online self-service services, and give the system as much as possible to the part that can be completed by the system and the user's autonomous operation, avoiding excessive dependence on manual labor; at the same time, try to give the business or operation that cannot be completely handed over to the machine to be performed. Use man-machine coordination to reduce manual work;
• The financial securities industry needs to establish a customer data center to provide a unified data service platform for customer operations and customer services;
• The financial and securities industry needs to use big data and artificial intelligence methods to enhance the verification of customer identities, grasp the accurate intentions of customers, identify customer risk tolerance, and avoid business risks and operational risks.

(2) How to improve customer experience in the process of investment and financial management to help customers realize their investment expectations simply and happily?

• The financial and securities industry needs to integrate "investor education" into the details of the product to improve the level of investment decision-making in the scene; design a richer portfolio of products to provide investors with more investment options and stabilize investment risks; customer-centric , Establish cross-channel customer service and realize seamless connection of different channels.

In addition to the two major needs mentioned above, the financial securities industry is embracing the following four trends:

• The trend of online: financial technology development and operation centralized, automated, and intelligent;
• Retail institutionalization trend: the scale of public funds and asset management users has developed rapidly;
• Difficulty trend of compliance supervision: current user service capabilities do not match with requirements, user risk level identification and matching;
• Insufficient self-research capabilities of science and technology departments of securities institutions: the majority of products are purchased out of the market, and there are few personalized self-researches.

In order to cope with the above-mentioned needs and trends, the intelligent operation product system needs to cope with multiple scenarios, including three major scenarios: customer multi-channel service scenarios, customer private domain operation scenarios, and internal large-scale operation scenarios. In the core architecture of these scenarios, a large amount of structured and unstructured data will be used, as well as corresponding technical components, such as application technology architecture components such as Docker, Kubernetes, and CI/CD pipelines. Among them, the key component used to analyze unstructured data includes the Milvus vector database. The overall architecture of the three scenarios is shown in the following figure:

Scene 1: Intelligent customer service robot

Through the introduction of the open source vector database Milvus, Palm Data Technology has built a complete set of intelligent question answering system to help online customer service answer questions and recommendations.

The intelligent Q&A component can realize the automatic reply of 80% of common questions, greatly reducing the workload of manual customer service. In addition, the question answering system can not only complete simple question and answer, but also make comprehensive opinion recommendations. Based on a large number of requests sent by the user, the system will recall the most suitable answer for the user and recommend it to the user, implementing functions such as "fund product card knowledge" recommendation, "fund manager card knowledge" recommendation, and "user input automatic completion".

At the same time, based on the open-source vector database Milvus to build a knowledge base search engine, it can also build an intelligent knowledge base for internal manual customer service to assist relevant personnel in quickly responding to customer questions.

It should be noted that in the financial securities business, the basic database is different from other industries, and some industry knowledge needs to be pre-processed. For example, product naming usually has its own rules. Fund names such as "Tiantianying" and "Fortress" do not often appear in other corpora, so it is necessary to train based on the existing corpus to complete data preprocessing. In the implementation process, the high-performance retrieval of the open-source vector database Milvus can achieve millisecond-level responses and recall thousands of data sets, greatly reducing development costs and shortening the project cycle. The subsequent system further refines the algorithm and screens out 10 data as Response result. The implementation process behind the intelligent customer service robot is shown in the following figure:

Scene 2:

According to the requirements of the industry customer suitability management measures, as well as anti-fraud, anti-money laundering and other compliance control requirements, the identification and review of customer identity 1617bdb46d81f7 is a technical field that the financial securities industry must focus on. In the past, accounts and passwords were mostly used to identify customers, but the passwords were easy to leak. Therefore, generally increasing the security level will require additional devices such as mobile phone dynamic passwords or U disks to avoid password protection loopholes, and then expand to face recognition technology based on Face ID. However, in some scenes that are not suitable for face comparison, it is necessary to use voice recognition and voiceprint comparison technology as a supplement. Compared with other biological characteristics, the voice of voiceprint features is very convenient and natural, user acceptance is high, and the cost is low, no additional recording equipment is needed during the call, the algorithm complexity of voiceprint recognition and confirmation is low, and its dynamic characteristics It has an exclusive advantage in application security.

In the financial industry, the technical requirements for voiceprint applications are:

First of all, the accuracy should be high. the accuracy of the recognition algorithm to the audio data set that meets the recognition requirements, especially the 1:1 identity verification scene, the accuracy requirement is up to 99.5% or more; combined with the face or other technologies, the difficulty of the system being broken will become exponential The level rises.

Second, the performance is better. The customer identity confirmation process of the tens of millions of voiceprint database needs to be able to achieve a second-level response, otherwise the user experience will be very bad; the voiceprint database must be seamlessly expanded horizontally to cope with the larger-scale voiceprint platform volume; In the identification scene of VIP customers or blacklisted customers, it is necessary to be able to quickly find similar results within a few seconds to 1 minute.

Finally, the cost should be low. acquisition end uses ordinary telephone or computer microphone, and the server end can be an ordinary X86 server; no special server hardware is required, and no GPU or other special hardware card acceleration is required to realize the voiceprint library and voiceprint platform when inferring and identifying the scene. Build. The realization process of the voiceprint scene is shown in the figure below:

By introducing the open source vector database Milvus, Palm Data Technology builds and accumulates a customer voiceprint library to help customers open accounts online and business opening scenarios, providing vector storage for customer identity data, retrieval and comparison, and blacklist customer identification services. The implementation process of voiceprint retrieval is shown in the following figure. First, input the target voice, perform feature extraction and model training, and save the feature vector in the Milvus vector database. When feature comparison is required, quick extraction and comparison can be performed. 1:1 voiceprint comparison refers to confirming " are you ", which is used for personal authentication, personal identity authentication, and mobile client identity authentication; 1:N voiceprint recognition is used to answer " you Who is it? " question is used to find the identity of the target in the vector database and check the duplicate. The application of the Milvus database has helped the business reach the technical indicators in the financial field and achieved the optimization of high-precision performance.

In the customer service scenario, when handling customer return visits or other business acceptance, it is necessary to check whether the securities business service process is compliant and verify communication skills. Combining the above multiple technologies can further form an intelligent voice quality inspection solution. The implementation process is as follows:

summary

Based on the open-source vector database Milvus, Palm Data Technology has built a question and answer and recommendation system for intelligent customer service robots, as well as a confirmation and recognition system in the voiceprint scene, which the relevant technical requirements with high accuracy, good performance, and low cost. We I hope that in the future, the Milvus vector database will support richer functions, adapt to a wider range of application scenarios, and be more widely used in the financial industry.