Dimensioning your Speech Server

Dimensioning rules

The Verba Speech Server is a server role, which can be installed as a

  • standalone server, or
  • co-located with a Verba Media Repository (both Express and Enterprise Edition).

Here are a couple of rule of thumbs for dimensioning your Speech Server:

 Standalone ServerCo-located with a Verba Media Repository
(installed within the same Windows Server OS)

Number of CPU Cores

Provide one CPU core per 30 simultaneous audio calls,
but use minimum 2 cores.

(E.g. 240 simultaneous calls: 8 CPU cores)

Provide one CPU core per 20 simultaneous audio calls,
but use minimum 2 cores.

(E.g. 160 simultaneous calls: 8 CPU cores)

Memory usageMinimum 4 GB (8 GB or more is recommend)8 GB or more
Impact on disk capacityThe size of the speech search index is ~10-12 Mbyte / hour
(This is up to 2x the size of WAV files using the GSM codec, which means your storage requirements can be up to 3x higher when you are using WAV/GSM.)

You can estimate your speech index storage requirements using the Storage requirements (make sure you select the GSM audio codec).

Manage index storage impact by:

  • indexing only what is needed for search
  • deleting indexes that are not needed in the future
  • compromising search result confidence (not recommended - contact support, if you need help with this)

Reference test result

Here is a reference test result run on a standard Intel server.

Server configuration

  • 2.4 Ghz Quad Core Intel CPU
  • 4 GB RAM 1066mhz DDR3
  • Samsung hd502IJ 500GB (not SSD)

Data input

  • ~22 hours of bidirectional phone calls in WAV/PCM format (~2.5 Gbyte - 1322 minutes)

Results

  • Indexing ready in ~10 minutes - 133x realtime - CPU above 90% (tests on SSD disks showed higher performance)
  • Search ready in ~16 seconds - 4957x realtime - CPU at 10% due to IO bottleneck (tests on SSD disks showed much higher performance)

In our most optimized server scenarios, we have seen single server indexing performances up to 300x realtime and search results close to ~50,000x realtime.