iFLYTEK: Improving Speech Recognition with AI

iFLYTEK optimizes their cloud for artificial intelligence based on 2nd Generation Intel® Xeon® Scalable processors.

Executive Summary

As a Chinese and global front-runner in the intelligent speech and artificial intel­ligence (AI) industry, iFLYTEK Co., Ltd. (iFLYTEK), which was founded in 1999, has established a leading international position in such technologies as speech recognition and natural language processing and has captured an over 70% share of the Chinese speech technology market.1 Looking ahead to the future, iFLYTEK is currently planning to explore new AI applications in cognitive fields on the basis of its iFLYTEK Cloud. This platform builds upon iFLYTEK AI technology as its core engine, incorporates data and expert knowledge from industries, and will enable a more effective AI industry ecosystem.


iFLYTEK is keenly aware of the importance of cloud computing platforms to realize these strategic goals. They also know that the development of AI technologies such as deep learning not only requires the improvement of top-layer applications, but also support in the form of computing, data processing, and algorithm optimization from underlying platforms. Achieving this strategy will require iFLYTEK to build a long-term win-win relationship with a partner who has extensive integration capa­bilities in the aforementioned technologies. Another need is to find a more diver­sified and optimized hardware platform to pair with iFLYTEK's Cloud for artificial intelligence (hereafter “AI Cloud”) and its applications.

“Intel is not only an outstanding hardware product provider, but also a leading full stack solution provider in the field of artificial intelligence. They are a trustworthy, dependable, all-round partner for our artificial intelligence strategy, helping us effectively explore paths to future innovation in artificial intelligence.” — Mr. Zhijiang Zhang, vice president of iFLYTEK's Cloud Computing Research Institute


Given these two considerations, iFLYTEK and Intel have entered into an in-depth col­laborative relationship. Intel's AI technology portfolio includes multi-purpose, pur­pose-built, and customizable hardware platforms, takes both hardware support and software optimization into account, and covers both end and cloud solutions. Intel has not only provided iFLYTEK with advanced hardware products like Intel® Xeon® Scalable processors as the “core” of iFLYTEK's AI Cloud, but has also been striving to accelerate the migration and optimization of relevant code and models from graph­ics processing units (GPGPU) to the Intel Xeon Scalable processor platform.

The two parties have achieved extremely positive results thus far. For one thing, iFLYTEK has already successfully run its AI Cloud on Intel Xeon Scalable processors. In addition, iFLYTEK has migrated large numbers of existing GPGPU-based AI applications to platforms based on Intel Xeon Scalable processors and has witnessed excellent performance optimization results. These underscore the first steps both parties have achieved in exploring pathways and directions for advance­ment in the future development of AI.

Data acquisition, development of algorithms, and computing power all impact AI workload performance. Until now, iFLYTEK has met the first two requirements pretty well. However, there are still enormous computing power challenges to overcome.

In order to enhance deep learning performance, iFLYTEK's “Super Brain Project” seeks to emulate human brain neurons in order to give the company's intelligent speech devices rudimentary human thinking capabilities. To achieve this goal iFLYTEK must overcome the immense challenge of process­ing vast quantities of training data and model parameters. To realize this deep simulation requires an even larger-scale supercomputer cluster, better deep learning algorithms as well as customized and dedicated neural network chip systems.

iFLYTEK is therefore focusing on Intel—another AI industry leader. Mr. Zhijiang Zhang—vice president of the iFLYTEK's Cloud Computing Research Institute—describes the coop­eration between the two parties as follows: “Our partnership with Intel is for the purpose of exploring future developmen­tal pathways and directions in the field of artificial intelli­gence. We are not merely looking for a hardware partner in the conventional sense.”

Strategic Cooperation: Starting with “Finding an Alternative Solution”

Enabling computers to “understand” human speech is the key task of intelligent recognition. After achieving widespread acclaim for its speech recognition achievements in recent years, iFLYTEK has not only established robust speech recog­nition capabilities with speech recognition systems based on DNN (Deep Neural Networks), RNN (Recurrent Neural Network), and RNN-CTC (RNN with Connectionist Temporal Classifica­tion) algorithms, but has also developed several innovative intelligent speech recognition frameworks, such as FSMN (Feedforward Sequential Memory Neural Networks) and DFCNN (Deep Fully Convolutional Neural Network), which are leading the advancement of speech recognition technology.

“The 2nd Generation Intel Xeon Scalable processor with integrated Intel Deep Learning Boost (VNNI) technology, will greatly boost our AI computing. It will also improve the TCO of our AI Cloud by allowing us to remove the previous GPU card. In real workload testing in AI cloud with 2nd Gen Intel Xeon Scalable processors, we can get similar or better performance with VNNI enabled compared to the GPU solution.” — Mr. Zhijiang Zhang

To further refine speech recognition, massive data training must be conducted on these frameworks, which in turn will bring enormous workloads for computing. The deep learning infrastructure adopted in speech recognition is about linking computing resources to a parallel file system over high-speed networks, on which the compute engine—the core of speech recognition—is developed and used in various types of model training and computing. It's obvious that paral­lel computing power and high-speed network transmission capacity will directly affect the operating performance of this infrastructure.

Although iFLYTEK previously employed conventional GPGPU solutions, the impressive potential of 2nd Gen Intel Xeon Scalable processors in this type of application has made iFLYTEK's engineers eager to work with it. The engineers plan to test the performance of the processor in deep learning systems.

“The 2nd Generation Intel Xeon Scalable processor with integrated Intel Deep Learning Boost (VNNI) technology, will greatly boost our AI computing,” Zhang explained. “It will also improve the TCO of our AI Cloud by allowing us to remove the previous GPU card. In real workload testing in AI cloud with 2nd Gen Intel Xeon Scalable processors, we can get similar or better performance with VNNI enabled com­pared to the GPU solution.”

“We built a hot data cache with Intel® Optane™ SSDs in the AI Cloud to provide fast access for the models during comput­ing, which improve the average response time of AI Cloud significantly”, Zhang continued. “And we optimized the AI cloud architecture for the 2nd Gen Intel Xeon Scalable proces­sor and Intel Optane SSD combination so that we could not only achieve the best performance for each of them, but also balance the performance from the system level.”


The performance of the 2nd Gen Intel Xeon Scalable proces­sor and Intel Optane SSDs has not disappointed iFLYTEK's experts: In terms of performance, Intel Xeon Scalable proces­sor can integrate up to 28 high-performance cores and the greater vector width obtained with Intel® Advanced Vector Extensions 512 (Intel® AVX-512) technology makes it espe­cially well-suited to high-load parallel computing situations. At the same time, this processor also offers high scalability and reliability under high-performance workloads, making it suitable for the rapid reasoning of complex neural networks in deep learning.

“The 2nd Generation Intel® Xeon® Scalable processor with integrated Intel® Deep Learning Boost (Intel® DL Boost) (VNNI) technology, will greatly boost our AI computing.” — Mr. Zhijiang Zhang


2nd Gen Intel Xeon Scalable processors and Intel Optane SSDs together can even better meet the need of the iFLYTEK AI Cloud to accelerate diverse applications. AI Cloud typically runs multiple applications programs and these programs have different hardware requirements. While a deep neural network will need high parallel floating point computing capacity, many other applications do not have these needs and may make frequent use of processing units that are suited to general purpose computing.

Accordingly, to adapt to different applications, the AI Cloud needed to be configured with different hardware. However, growing the diversity in hardware platforms would increase the complexity of purchasing, deployment, operation, and maintenance, which will in turn lead to much higher costs. As a consequence, to address complex application needs, the ideal solution is to select an integrated hardware platform that is able to simultaneously perform the acceleration of general tasks and AI applications. 2nd Gen Intel Xeon Scal­able processor is the choice because it can adapt to different application loads, help increase the configuration flexibility of AI Cloud, and deliver better scalability while reducing the platform's complexity and cutting costs. In addition, when applications originally running on GPGPUs are migrated to 2nd Gen Intel Xeon Scalable processor and optimized, it is also possible to reduce total cost of ownership (TCO) while ensur­ing high computing performance (HPC).

Migration of Applications: In-Depth Cooperation Between the Two Parties' Engineering Teams

In the collaboration between the two parties, Intel and iFLYTEK were aware that: The achievement of more in-depth cooperation as well as better application migration and optimization results will require extend the cooperation to the code level. Only in this way, can they attain the highest degree of optimization.

As contact and communication between the two parties increases, both parties' technical experts have been estab­lishing highly effective communication, interchange, and cooperation mechanisms in such areas as code migration and optimization as well as hardware optimization.

First, in the area of code migration and optimization, the two parties have embarked on the following tasks:

  • Clarifying roles and responsibilities. To respect and protect iFLYTEK's intellectual property rights, iFLYTEK abstractly summarized its deep neural network model and matrix scale etc., and then provided program code and data to Intel; Then, Intel migrated the codes and deeply optimized them based on features of Intel Xeon Scalable processor, such as high-efficiency cores, memory usage, and ultra-wide vector width. This in turn ensures that the performance of applications based on Intel Xeon Scalable processor meets iFLYTEK's expectations. Afterwards, the results of optimization, including code and performance parameters, will be sent back to iFLYTEK, allowing iFLYTEK to realize the optimized solution in its actual environment, perform debugging, and testing, verify the accuracy and effectiveness of optimization and ultimately help boost the performance of applications in actual production environments. Furthermore, iFLYTEK can apply its optimization methods and experience to the migration and optimization of other AI applications.
  • Exchange ideas and work experiences regularly. Both parties' technical experts hold bi-weekly routine telephone conferences. In these conferences, the two parties report on their work progress, engage in interchange, and discussion concerning technical issues, resolve any problems that have been encountered, determine optimization directions and plan next moves. In addition, the two parties have also held occasional mutual visits and meetings of high-level personnel and arranged their technical experts to have face-to-face contact to upgrade and maintain the test environment of iFLYTEK.

With regard to hardware optimization, in order to enhance the storage performance of the AI Cloud, iFLYTEK also adopted Intel Optane SSDs, which possess Intel® Optane™ technology offering a combination of high throughput, low latency, high quality of service (QoS), and high endurance. The Intel® SSD technology team has helped iFLYTEK to fully optimize the performance of Intel Optane SSD in its AI Cloud and thereby transcend existing storage bottlenecks and tap the full potential of Intel Xeon Scalable processor.

Intel has also provided many reference suggestions concern­ing the architecture of iFLYTEK's cloud computing platform. In-depth interchanges were arranged between iFLYTEK's technical team and the Intel's cloud computing team and the big data technology team. For iFLYTEK's AI Cloud, Intel has provided all-round consulting services and gave sugges­tions concerning such aspects as the design of the platform's underlying architecture, optimization of Cloud software and big data software, and operational maintenance and man­agement of the Cloud etc., helping iFLYTEK to successfully launch the AI Cloud.

Through full-scale and in-depth collaboration, Intel and iFLYTEK have achieved very satisfactory outcomes. iFLYTEK was keenly aware of Intel's full stack solution capacity in the areas of AI and cloud computing. In particular, iFLYTEK has migrated its AI applications formerly on GPGPU to Intel Xeon Scalable processor, made full use of the computing capabili­ties obtained from optimization and simplified the complexity of AI Cloud deployment and maintenance, all while reducing TCO. Throughout this process, Intel has had a deep under­standing of the AI leader iFLYTEK's application optimization experience, views, and core needs. Many AI-related software tools and libraries of Intel, such as Intel® Parallel Studio and Intel® Math Kernel Library for Deep Neural Networks (Intel® MKL-DNN), have been significantly improved due to this col­laborative effort, enabling both parties to reap the benefits of a win-win partnership.

Future: Standing Together at the Forefront of Artificial Intelligence

The foregoing series of in-depth collaboration has enabled Intel and iFLYTEK to gain a better understanding of each other's AI capabilities and visions. They were also conscious of the need to change their thinking from conventional hardware cooperation to in-depth collaboration at both com­mercial and strategic levels. This is an important opportunity for the two companies to embrace the future and ride the AI wave. With this background in mind, iFLYTEK has already begun testing and using more products and technologies provided by Intel. In the iFLYTEK AI Cloud provided to users and developers in various fields and industries, iFLYTEK has already begun adopting the all-new 2nd Gen Intel Xeon Scal­able processor, Intel Optane SSD, and Intel® FPGA which can be customized and offers flexible support for applications.

Looking ahead, Intel and iFLYTEK will continue to deepen technological collaboration on the basis of the framework specified in the two parties' strategic MOU and make joint efforts in the areas of market research and formulation of market strategies.

Explore Related Intel® Products

Intel® Xeon® Scalable Processors

Drive actionable insight, count on hardware-based security, and deploy dynamic service delivery with Intel® Xeon® Scalable processors.

Learn more

Intel® Deep Learning Boost

Intel® Xeon® Scalable processors take embedded AI performance to the next level with Intel® Deep Learning Boost (Intel® DL Boost).

Learn more

Intel® Optane™ DC SSDs

Intel® Optane™ technology is the first major memory and storage breakthrough in 25 years.

Learn more


英特尔® 技术的特性和优势取决于系统配置,并可能需要支持的硬件、软件或服务激活。实际性能可能因系统配置的不同而有所差异。没有任何计算机系统能够保证绝对安全。请咨询您的系统制造商或零售商,也可登录 www.intel.cn 获取更多信息。// 性能测试中使用的软件和工作负载仅在英特尔® 微处理器上针对性能进行了优化。SYSmark 和 MobileMark 等性能测试使用特定的计算机系统、组件、软件、操作和功能进行测量。上述任何要素的变动都有可能导致测试结果的变化。您应该查询其他信息和性能测试,以帮助您对正在考虑购买的产品作出全面的评估,包括该产品在与其他产品结合使用时的性能表现。如欲了解更多完整信息,请访问 www.intel.cn/benchmarks。// 性能结果基于配置中所规定日期的测试,可能无法反映所有公开的安全更新。有关详细信息,请参见配置信息披露。没有任何产品或组件能保证绝对安全。// 所描述的成本降低方案仅用作示例,表明某些基于英特尔® 的产品在特定环境和配置下会如何影响未来的成本,并节约成本。环境各不相同。英特尔不保证任何成本和成本的节约。// 英特尔并不控制或审核本文档引用的第三方基准资料或网站。您应访问引用的网站,确认参考资料准确无误。// 在某些测试案例中,结果以英特尔内部分析或架构模拟或建模为基础来评测或模拟,且仅供参考。您的系统硬件、软件或配置的任何不同均可能会影响实际性能。