"Hello, is this Mr. Li Li?" Maybe many people found out that it was a robot customer service after dialing the customer service number, and hoped to switch to manual service as soon as possible. But SF's "customer service robot" is "out of the ordinary".
SF Express has become a leading comprehensive express logistics service provider in China and the fourth largest express delivery company in the world. Relying on its leading technology research and development capabilities, SF Express provides users with convenient, professional and warm delivery services. As the smart brain of SF Express, SF Technology focuses on realizing the digital transformation and upgrading of the logistics network and supply chain chassis, and helps improve customer experience by opening up the digital closed loop of operations, sales, experience and other links and sectors.
The natural language processing (NLP) team of SF Technology has developed a speech synthesis system of "smart outbound robot" based on the Paddle Speech speech model library , which will be officially launched on the eve of Double 11 in 2022 and will continue to operate stably, helping to improve business quality synergy.
It is reported that Li Li, the chief NLP engineer of SF Technology, led the team to complete the development, testing and online application of the speech synthesis self-developed function in less than half a year with the help of the open source speech model library provided by the Flying Paddle deep learning platform . SF Express’ national customer service hotline “95338” provides monthly outbound call notifications, return visits, research, etc., among which millions of calls are completed through the “smart outbound robot”.
For long-term development, self-research is for better service
SF's courier brother usually communicates with customers by phone through "95338" before delivering them to the door, asking if the customer is at home, whether they can deliver now, or when they will be delivered. The "smart outbound robot" solution is to use an outbound robot to communicate with customers in advance instead of the courier brother making a phone call. This solution has already been widely used, helping couriers to greatly improve work efficiency, and also for users. Bring better service.
Previously, SF Express launched the "smart outbound robot" service through external procurement of speech synthesis technology and products. However, with the rapid growth of SF Group's business volume in recent years, when the monthly traffic volume reaches one million levels, the externally purchased products cannot meet the current business needs.
On the one hand, external suppliers can only provide some basic services, and it is difficult to achieve differentiated customization for SF Express, and the business innovation and improvement cycle is long, which cannot meet the needs of continuously improving customer service experience; on the other hand, the company expects to In the field of construction, we will build a closed-loop of self-developed technology in the whole chain to more flexibly meet the targeted needs of customers and follow the wave of artificial intelligence development. In this context, SF Technology decided to create a self-developed "smart outbound robot".
Significant results have been achieved in improving quality and efficiency, employees are willing to use it, and customers are more satisfied
As the technical leader of SF Technology's self-developed "Smart Outbound Call Robot" project team, Li Li introduced that two important parts of the self-developed overall intelligent outbound call system are ASR (Automatic Speech Recognition) automatic speech recognition technology and TTS (Text-To-Speech) speech synthesis technology.
"Now is not the time to build wheels from scratch, and it is impossible to start from scratch with self-development. There are also open source speech synthesis systems abroad, but most of them are in English, and **Flying Paddle Speech Model Library Paddle Speech supports Chinese better* *, and we also learned that Paddle Speech has a very mature solution.” At the end of April 2022, after conducting comparative research on the deep learning platform and model library in the industry, Li Li began to lead the team to try.
The paddle speech model library Paddle Speech provides an end-to-end development kit and rich tool components, allowing SF Technology to develop and implement applications faster and better like building blocks. "During the project implementation, the team sought help from the Flying Paddle community many times, and received positive responses from the Flying Paddle 's professional R&D team and operation team, clearing the way for the project to progress in a timely manner, which allowed our team to efficiently complete the speech synthesis system The development and various testing work." Li Li concluded.
Speech synthesis system, using the front end, acoustic model, and vocoder to process the input text, and then output the synthesized voice data, which is then used in SF Express’s intelligent outbound call robot to interact with the called party by voice, and conduct multiple TTS in real time for each call convert.
Schematic diagram of the speech generation system
Specifically, based on the Paddle Speech speech model library , Li Li led the team to further train two algorithm models, FastSpeech 2 and HifiGAN, to complete speech synthesis tasks. The front-end processing module processes the input text into phoneme ids, and the acoustic model selects the non-autoregressive algorithm model FastSpeech 2, which significantly improves the expressiveness of the synthesized speech , and the non-autoregressive model is faster ; the vocoder selects a GAN-based framework The HifiGAN algorithm model is not only fast, but also produces better sound quality. In actual use, it meets the real-time rate requirements of robot intelligent interaction . Paddle Speech, a speech model library, makes the effect of text-to-speech synthesis even better .
SF Express' self-developed "smart outbound call robot" has achieved a large-scale application of one million calls per month, and the actual application effect can also be reflected from multiple aspects. For employees, more and more courier brothers are willing to use one-click outbound calls and let robots do their work for them. The customer satisfaction of SF's internal evaluation and self-developed system has also continued to improve, and the system has achieved remarkable results in improving the quality and efficiency of the system.
Create the ultimate customer experience and look forward to more room for development in the future
Li Li said frankly, "'Service experience' is one of SF's core competitiveness, which has been valued for a long time. We even call it the 'life' of the express logistics industry. Continuously improving digital technology capabilities to empower internal operations can help SF Express Provide customers with more personalized and warm services.”
For example, we need to correct a large number of polyphonic characters, provide voices with thousands of faces, and support blessings or messages generated by the customer's own voice. "Meeting more personalized needs and achieving the ultimate customer experience is difficult for suppliers that provide standard services, or it is a business that suppliers are unwilling to do. From a strategic perspective, with this self-developed voice After the synthesis technology, we will have more opportunities to make personalized and warm products in the future. The logistics industry scene combined with more AI technology applications will have more possibilities and greater room for imagination.”
At the same time, relying on technology products and data assets that have been accumulated for many years, SF Technology quickly empowers external customers in combination with the characteristic scenarios of various industries: through digital and intelligent means, it embeds in the business value chain to help customers transform and upgrade their supply chains. SF Technology is committed to building a smart supply chain ecosystem in the digital age and becoming a leader in the smart supply chain.
Baidu Flying Paddle is committed to enabling developers and enterprises to develop AI applications more quickly and conveniently, and as a basic platform to support all walks of life to achieve industrial intelligent upgrading. Li Li also looks forward to using the Flying Paddle deep learning platform to expand broader application scenarios in the future.
Robot customer service (Baidu Wenxin Yige AI painting generation, author eagle_kang)
reference
- Paddle Speech project address
https://github.com/PaddlePaddle/PaddleSpeech
- Application example video: Generating a virtual human voice using Paddle Speech's speech synthesis module
https://paddlespeech.readthedocs.io/en/latest/tts_demo_video.html