Monitoring SpeechKit Hybrid services
Written by
Updated at August 25, 2023
SpeechKit Hybrid collects and stores metrics in Prometheus format. For more information about metric types, see the Prometheus documentation.
SpeechKit Hybrid service metrics are available at the URL:
<IP address>:17002/metrics/prometheus
Where:
<IP address>: IP address of the SpeechKit Hybrid service in your network.
Common SpeechKit Hybrid metrics
| Metric | Type | Description |
|---|---|---|
grpc_code_* |
COUNTER | Number of responses with a certain gRPC status code. Each code refers to a separate metric: grpc_code_ok, grpc_code_canceled, etc. For a list of codes, see the gRPC documentation |
Speech synthesis metrics
| Metric | Type | Description |
|---|---|---|
synthesis_sps |
GAUGE | Number of seconds of synthesized text generated in a second at SpeechKit Hybrid runtime |
synthesis_latency |
HISTOGRAM | Service response time histogram in milliseconds |
Speech recognition metrics
| Metric | Type | Description |
|---|---|---|
in_flight_requests |
GAUGE | Number of active sessions |
recognize_sps |
GAUGE | Number of seconds of recognized speech processed in a second at SpeechKit Hybrid runtime |
recognize_rtf |
HISTOGRAM | Distribution histogram of a Real Time Factor value that is equal to the ratio of recognition time to the duration of recognized speech expressed as a percentage |
Performance monitoring
To monitor SpeechKit Hybrid performance, set the target metric values:
synthesis_sps: Required number of seconds of synthesized text generated in a second.recognize_sps: Required number of milliseconds of recognized speech generated in a second.synthesis_latency: Allowed speech synthesis app response delay.recognize_rtf: Allowed ratio of the time it takes to recognize text to the length of the text itself expressed as a percentage. The recommended value is100, as a larger value means that a server cannot handle the stream of requests.
Set up metric monitoring to track their deviation for the worse from the set target values:
- Decrease for SPS metrics.
- Excess for other metrics.