The Intelligent Voice system is a batch processing system designed to process hours of audio recordings as efficiently as possible. We also offer systems designed for real-time transcription, low latency key word spotting, IoT hub devices and embedded applications and more - for details on running these solutions on AWS please contact us.
Minimum Requirements
A minimal installation of Intelligent Voice 5.2 with GPU acceleration on AWS requires a single GPU instance type with 32GB RAM and 4 CPUs.
As of Jan 2022, the g4dn.2xlarge or type is recommended. Other supported sizes:
- g4dn.2xlarge or larger g4 types
- any p3 type
- p4d.24xlarge
The minimum storage requirement is 500GB and minimum local storage (for swap file) is 110GB.
This configuration will process up to 1000 hours per day.
Production deployment examples
Small, medium and large examples of full production system deployment with autoscaling.
These system uses a single application server, with multiple autoscaling groups based on images captured as a AMIs. The current state of the gearman queue is sent to CloudWatch and scaling rules created based on the number of jobs.
Small 1000-10,000 hours per day
This configuration was benchmarked in Jan '22 with 1000 hours per day: Benchmark results. The approximate cost per audio hour processed in the benchmark was USD $0.04. The cost per hour is lower with more data.
Application server
Always on, includes MariaDB database
m5a.4xlarge
ASRWorker
spot instance autoscaling from AMI
g4dn.xlarge with 180GB storage (swap file 110GB on local storage)
VADWorker
spot instance autoscaling from AMI
c5.xlarge with 50GB storage
TaggerWorker
spot instance autoscaling from AMI
c5.xlarge with 50GB storage
Medium 10,000 - 50,000 hours per day
The target cost per audio hour at 10,000 audio hours per day is $0.02
Application server
Always on
c5d.8xlarge
Database
Amazon RDS for MariaDB
db.m6g.2xlarge
Triton inference server
always on
c5.xlarge with 50GB storage
ASRWorker
spot instance autoscaling from AMI
g4dn.xlarge with 180GB storage (swap file 110GB on local storage)
VADWorker
spot instance autoscaling from AMI
c5.xlarge with 50GB storage
TaggerWorker
spot instance autoscaling from AMI
c5.xlarge with 50GB storage
Large 50,000-250,000 hours per day
The target cost per audio hour at 10,000 audio hours per day is $0.01
Application servers
Always on
c5d.8xlarge
Database
Amazon RDS for MariaDB
db.m6g.2xlarge
Search nodes
Running sphinxsearch, autoscaling based on CPU usage, approx 1 node per 50k audio hours per day.
r5.2xlarge
Triton inference server
always on
g4dn.xlarge with 180GB storage (swap file 110GB on local storage)
ASRWorker
spot instance autoscaling from AMI
g4dn.xlarge with 180GB storage (swap file 110GB on local storage)
VADWorker
spot instance autoscaling from AMI
c5.xlarge with 50GB storage
TaggerWorker
spot instance autoscaling from AMI
c5.xlarge with 50GB storage
Configuration recommendations
Recommended minimum partition sizes for Red Hat app server, allocating additional space to /var and /tmp for higher performance:
Filesystem Size Mounted on
/dev/mapper/rootvg-rootlv 40G /
/dev/mapper/rootvg-usrlv 28G /usr
/dev/sda2 494M /boot
/dev/mapper/rootvg-optlv 190G /opt
/dev/sda1 500M /boot/efi
/dev/mapper/rootvg-homelv 6.0G /home
/dev/mapper/rootvg-varlv 150G /var
/dev/mapper/rootvg-tmplv 35G /tmp
(tomcat8 temp files are in default location /opt/apache-tomcat-8.5.61/temp/)
and proc node:
Filesystem Size Mounted on
/dev/mapper/rootvg-rootlv 40G /
/dev/mapper/rootvg-usrlv 20G /usr
/dev/sda2 494M /boot
/dev/mapper/rootvg-varlv 310G /var
/dev/mapper/rootvg-tmplv 25G /tmp
/dev/sda1 500M /boot/efi
/dev/mapper/rootvg-homelv 6.0G /home
/dev/mapper/rootvg-optlv 20G /opt
Comments
0 comments
Please sign in to leave a comment.