System Requirements

Before installing the CPQD ASR, please make sure your system meets the requirements described in this section.

Operating System

Linux Red Hat 7.8 or CentOS 7.8, 64 bits.

Use the yum update to upgrade the CentOS version.

To verify your installed version: cat /etc/redhat-release.

Processor

The processor must be one of the following models:

  • Intel® Xeon Phi™ processor

  • Intel® Xeon® processor

  • Intel® Core™ processor family

  • Intel Atom® processor

If you use processor that is not listed here, performance may suffer, and even require a different installer. In this case, please contact CPQD.

The number of simultaneous recognitions a machine running CPQD ASR is able to attend is directly associated to the number of vCPUs and the processing speed of each vCPU.

The reference machine used has 8 vCPUs @ 2.9 GHz and is able to handle:

  • Free speech: 30 simultaneous recognitions.

  • Grammar (50 items): 60 simultaneous recognitions.

This performance may vary, since it is affected by the characteristics of the processor, audio and grammar.

The machine’s capacity can be increased by increasing the number of vCPUs. Generally speaking, the number of simultaneous recognitions is proportional to the machine’s CPUs, but with losses. In tests with 8 and 18 vCPUS, a loss of around 10% was observed.

Note

The performance of speech recognition on machines with the recommended number of CPUs and clock speed can vary, depending on other characteristics of the processor, such as how fast it communicates with the memory. In fact, a much poorer performance was observed in processors launched prior to 2012.

To insure this performance, the CPQD ASR instance must use all the vCPUs of the same processor socket. For example, a machine with 32 vCPUs distributed in two sockets with 16 vCPUs each should be able to run two instances of the CPQD ASR server (through a virtual machine, Docker, etc.), each one using 16 vCPUs of just one socket.

Memory

For free speech recognition,the machine must have at least 8 GB of RAM memory, supposing there would be 6 GB of free memory for running the ASR. If the application will use only grammars and does not foresee the use of free speech, a machine with 4 GB of RAM memory is enough.

Disk

To install the software and models and collect logs, we recommend a machine with at least 30 GB of disk free space.

Number of Machines

To define how many machines running ASR are needed to serve a given application, the capacity of the machine and the requirements of the application must be taken into account, in other words, how many simultaneous recognitions the application will need to perform and how many simultaneous recognitions each machine is able to run.

Once in production, the configured system must be monitored to define if more machines running ASR are needed, or if it has enough power to handle a larger number of requests.