Qiying Tailun: AI voice chip three-step development, continuous iteration of offline NLP functions

On May 17, at the "14th Songshan Lake China IC Innovation Summit Forum" hosted by VeriSilicon, Wang Shujuan, chip development director of Chengdu Qiyingtailun Technology Co., Ltd., said that Qiyingtailun's edge AI chip products are currently being used in industry-leading home appliances and various smart terminals such as Haier, Gree, Midea, and AUX.

Founded in 2015, Qiyingtailun is a full-stack technology solution provider integrating voice chips, voice algorithms, development platforms, and application solutions. The following year, the company successfully launched its first end-side AI voice chip C11006. After 9 years of development, it has formed 4 major series and 20 chip models, covering a series of product layouts including end-side AI voice chips, AI voice Wi-Fi Combo, Al voice BLE, etc.

Specifically in terms of products, Qiying Tailun has gone through three major iterations and eight minor iterations, marking a new height in its technological evolution. Back in 2017, the first version of the technology was fully launched on the market, which realized multiple functions such as single-point intelligence of the device, end-side voice semantic recognition, and AI recognition, bringing users an unprecedented intelligent experience.

"Nowadays, many people pay great attention to online security issues, so offline application solutions are very suitable for now. There are many terminal-side TTS applications, such as hospital announcements, bus station announcements, ETC equipment broadcasts, etc." said Wang Shujuan.

Therefore, the offline NLP market has developed very rapidly. Wang Shujuan divides the development of Qiyingtailun into three stages. The first stage is the layout of natural language AI chips, solving the problem of interactive entrances from scratch, from command words to the understanding of general natural language. The second stage allows service robots to enter thousands of households, solve the problem of better service, integrate multi-sensory capabilities such as voice, touch, and vision, better perceive user intentions, and combine various perceptions to make intelligent decisions and provide more complete services; the third stage is to let everyone have a guardian elf to achieve cross-terminal and cross-time and space interaction. The guardian elf with private customization attributes can become a comprehensive life manager, health and safety guard, encyclopedia knowledge mentor, and a confidant friend who accompanies the soul.

If the first phase solved the problem of creating something from scratch, the second phase focused on introducing unique personalized recognition technologies, such as dialect adaptation, to solve the usage problems caused by differences in different regions and dialects. This innovative technology enables the device to autonomously learn and adapt to various dialects, ensuring that a wide range of user groups, including the elderly, can easily use smart terminal devices.

Since 2021, the company has entered the third stage of development. In this stage, Qiyingtailun launches new chips every year. At the "14th Songshan Lake China IC Innovation Summit Forum", Qiyingtailun officially released the CI135X series chips, which include five models. The core goal of this stage is to solve the problem of offline NLP recognition with richer natural semantics, aiming to provide users with more intelligent and convenient services.

CI135X uses the patented 3.5th generation brain neural network processor core (BNPU V3.5) of Qiyingtailun Company, with a maximum main frequency of 210MHz, and has stronger and more flexible computing performance, supporting neural networks and vector parallel operations such as DNNTDNNRNNCNNLSTM. The chip supports offline speech recognition, deep learning noise reduction, local self-learning, voiceprint recognition and other functions; it has stronger anti-noise ability, and the recognition rate exceeds 98% in quiet. The chip integrates PMU power management unit, high-precision RC oscillator, single-channel AD, DA and common MCU general interfaces, etc., and only a small number of peripheral devices such as resistors and capacitors are needed to develop various intelligent voice product hardware solutions. It is extremely cost-effective and can meet various speech recognition applications.

The basic application scenarios of the 130X series are now in the mid-to-high-end application market. Among them, the 135X is designed for small appliances, lighting, toys and other application scenarios that only require simple instructions, providing the most cost-effective solution. The SOP8 package only requires 3 capacitors. In addition, 1361 and 2361 meet the needs of small programs, and 1392 is suitable for scenarios with strict size requirements.

According to the introduction, the CI135X series has three characteristics. First, it has low cost, a chip-as-solution, a very simple BOM, high integration, and streamlined peripherals. In addition to the MIC and speakers, only a few resistors and capacitors are required at the board level. Second, it has powerful natural language and interactive capabilities. The recognition rate in a quiet environment is 98.96%, and under 70dB noise, the recognition rate is 85% and the wake-up rate is 90% (noise reduction mode). BNPUV3.5 has strong computing power and supports 300 entries. Third, it has three packaging specifications of SOP8/SOP16/OFEN32 and five models to meet different applications and patch requirements.

CI135X has a very high integration level, with built-in crystal oscillator, power management, etc., thus reducing BOM

C135X Block Diagram