Assist Elephant Sound Technology in achieving AI audio algorithm simulation calculation acceleration, greatly improving the efficiency of intelligent noise reduction research and development!
Elevoc was founded in Shenzhen in 2017 and is a leading global enterprise in machine auditory artificial intelligence, focusing on the research and development of intelligent speech enhancement and interaction solutions. Based on the theory of Computational Auditory Scene Analysis (CASA) and deep learning technology (DL), the company provides solutions for smartphones PC、 We provide cutting-edge voice technology in the fields of wearable devices, smart cars, smart homes, and VoIP cloud communication, committed to creating a natural and efficient human-computer interaction experience.
System stability
7 × 24-hour full load operation with zero downtime, temperature fluctuation of ± 2 ℃, and 99% availability guarantee.
Processing efficiency
The end-to-end delay of 192kHz audio stream has been reduced from 153ms to 35ms, resulting in a 67% increase in single node concurrency,
Project lifecycle
The project development cycle has been shortened by several weeks compared to the past, which has advanced the project progress
Project Background
Elevoc was founded in Shenzhen in 2017 and is a leading global enterprise in machine auditory artificial intelligence, focusing on the research and development of intelligent speech enhancement and interaction solutions. Based on the theory of Computational Auditory Scene Analysis (CASA) and deep learning technology (DL), the company provides solutions for smartphones PC、 We provide cutting-edge voice technology in the fields of wearable devices, smart cars, smart homes, and VoIP cloud communication, committed to creating a natural and efficient human-computer interaction experience.
Challenge pain points
The bottleneck of computing power and video memory restricts real-time processing performance
The insufficient computing power and graphics memory of the original hardware resulted in high latency when processing 192kHz voiceprints, which could not meet real-time interaction requirements and also forced feature shard loading.
Insufficient scalability and multi card collaboration efficiency
The original node has a maximum of 4 GPUs, which limits the linear expansion of computing power. The scalability of the four cards cannot support large-scale voiceprint model training, making it difficult to achieve collaborative offloading of computation and storage in terms of compatibility.
System reliability and operational risks under high load
GPU temperature>85 ° C causes frequency reduction, non redundant power supply poses a risk of data loss in voltage sensitive scenarios, and traditional maintenance mode requires shutdown to replace components, affecting business continuity.
Solution
Computing power upgrade and energy efficiency optimization plan
Deploy the Shuju Hongxin HG8380S server, equipped with 10 NVIDIA A800 graphics cards and 800GB of large capacity video memory, to solve the high concurrency requirements for speech model training and inference.
Data processing latency and future compatibility solutions
Supports NVMe/SATA/SAS multi protocol hybrid access, with a total of 24 hot swappable disk slots. From structured speech data collection to concurrent writing of training samples, data throughput is unobstructed and scalability is worry free.
High reliability power supply and intelligent cooling solution
By using 8 hot swappable fans combined with structured cooling of the chassis and 4 × 2000W titanium redundant power supply, 99.99% availability is achieved to ensure stable operation under high loads.
Project Benefits
Relying on a team of top machine auditory scientists and deep learning technology, we provide advanced speech enhancement and interaction solutions for various intelligent devices, promoting the development of human-computer interaction experience towards a more natural direction. When faced with research and development challenges, the company upgraded to the Shuju Hongxin HG8380S ten card GPU server based on the INTEL dual channel third-generation Xeon platform, equipped with 10 NVIDIA A800 graphics cards, significantly improving the computing power and efficiency of audio processing. This technological innovation not only accelerates project progress and reduces team pressure, but also lays a solid foundation for the company's leading position and future development in the field of intelligent voice.
Customer reviews
The intelligent voice solution has significantly improved our audio processing efficiency, and its professional team and leading technology provide strong support for the rapid landing of products
