Method
- Evaluate the load on both the Biometrics API and the votg-adapter service during concurrent enrollment transactions, focusing on real-time audio streaming of calls.
- Utilize an internally developed tool named "Scenario Tester" to simulate the actual call stream through WebSocket connections, as conventional testing tools lack this capability.
- The tool employs a default enrollment period (20 sec), and an audio file is transmitted in small 2 KB chunks during the simulation.
Assumptions
- None
Prerequisites
- No other active users in the environment during the load tests.
- HA has been enabled for all services (except db-migrator).
- Enrollment duration is set to 20 seconds.
- Speaker Change Detection, Synthetic Speech Detection and Noise Detection modules enabled during load test.
- Blacklist identification is excluded from load test scenario.
- Voiceprint Adaptation is excluded from load test scenario, since it's simply a recreation of the voiceprint.
Tests
- Load tests are conducted by simulating concurrent calls with using scenario tester.
- Scenario tester streams audio data to Knovvu Biometrics like real call scenarios and triggers enrollment transactions for each call concurrently.
- Hardware sizing for the certain number of concurrent enrollment transactions is calculated
- These test are conducted repeatedly to observe the stability of the services under the heavy load.
- The duration of the voiceprint creation operation after required amount of speech data collected is reviewed.
Result Format:
- Number of Successfull Scenarios
- Number of Failed Scenarios
- Detailed Timeline of each Failed Scenarios with logs
- Average duration of voiceprint creation
Notes:
- Speaker Change Detection, Synthetic Speech Detection and Noise Detection are the preliminary checks performed by the system before enrollment.
- Enrollment (voiceprint creation) is the heaviest operation done by the system.