Skip to main content
Rasa records OpenTelemetry metrics for monitoring conversations. This page describes each metric, when Rasa records it, and what its dimensions mean.

Configure metrics

Create an endpoints.yml file in your project and add a metrics section:
Replace the endpoint with your OpenTelemetry collector’s gRPC endpoint. Metrics can be enabled without configuring tracing.

Voice metrics

Voice metrics cover calls handled through streaming voice channels.

rasa.voice.call.duration

Records the time from the start of a streaming voice call until the call ends. Rasa records it once at the end of the call, so ongoing calls are not included. Use it to compare call durations and outcomes. Dimensions

rasa.voice.tts.time_to_first_audio

Records the time to the first audio chunk for one agent message. For a cache hit, timing starts when Rasa finds the cached audio. For a cache miss, it starts when Rasa sends the request to the TTS provider. The metric is not recorded if no audio is produced or TTS fails, even if fallback audio plays after the failure. In that case, the voice.latency.tts_first_byte_ms span attribute can still measure the fallback audio when tracing is enabled. Use it to compare time to first audio by provider and cache status. Dimensions

rasa.voice.tts.synthesis.duration

Records the time to complete TTS for one agent message. For a cache hit, timing starts when Rasa finds the cached audio and ends when cached-audio delivery completes. For a cache miss, it starts when Rasa sends the provider request and ends when audio generation completes. On failure, timing stops at the first failure and excludes fallback audio. If TTS neither completes nor fails, no measurement is recorded. Use it to compare TTS completion times by outcome, provider, and cache status. Dimensions

rasa.voice.playback.duration

Records the time from the first delivered audio chunk until playback completes, is interrupted, is cancelled, or fails. Rasa records it only after at least one audio chunk is delivered. If playback ends before delivering audio, this metric is not recorded, but rasa.voice.operation.failures can still count an unexpected failure. Use it to compare playback times by outcome. Dimensions

rasa.voice.operation.failures

Counts unexpected failures in streaming voice operations. Expected outcomes, such as an ignored barge-in, a DTMF timeout, or interrupted or cancelled playback, are not counted. Use it to identify failures by component, operation, provider, and error type. Dimensions

rasa.voice.barge_in.decisions

Counts one accepted or ignored barge-in decision for each transcript update Rasa evaluates, including partial transcripts. Decisions are made when interruptions are enabled and the caller speaks while agent audio is playing or waiting to play. One caller interruption can produce several ignored decisions before an accepted one. Other caller speech that does not result in a decision is not counted. Use it to compare accepted and ignored decisions and their reasons. Dimensions

rasa.voice.dtmf.decisions

Counts accepted DTMF input, keypresses received when Rasa is not waiting for DTMF input, and DTMF input attempts that time out. Keypresses are not counted individually while Rasa waits for the full input. Use it to compare keypad interaction outcomes without recording the digits entered. Dimensions

rasa.voice.silence_timeouts

Counts silence timers that expire and trigger a turn. Timers cancelled before they expire are not counted. Use it to compare counts across configured timeout values. Dimensions

rasa.voice.audio.bytes

Counts audio bytes processed during streaming voice calls. At the end of a call, Rasa records the total for each direction with more than zero bytes. Directions with no audio and ongoing calls are not included. Use it to compare incoming and outgoing audio volume. Dimensions

Turn metrics

Turn metric names apply across channels, but Rasa records them only for turns handled by Maestro agents. Supported inputs are text and, on streaming voice channels, speech, completed DTMF input, and accepted silence timeouts. Session start and end turns are not measured. A voice barge-in marks the current turn as interrupted instead of creating another measured turn.

rasa.turn.duration

Records how long Rasa processes a turn started by customer input. Rasa records it when the turn completes, is interrupted, is cancelled, or fails. Session start and end turns are not included. Use it to compare processing times and outcomes by input type. Dimensions

rasa.turn.time_to_first_output

Measures the time from when Rasa receives input for a turn until it produces the first output. It includes any delay before engine processing begins. For text turns, timing starts when the customer’s message arrives and ends when the first customer-facing text is produced. For voice turns, timing starts when Rasa receives the final speech transcript, DTMF input is complete, or a silence timeout is accepted. When consecutive final transcripts are merged, timing starts from the latest one. It ends when Rasa produces the first TTS audio chunk. Turns that produce no output for the customer in that channel are omitted, including turns interrupted before the first output. Use it to compare recorded input-to-first-output latency (an indicator of user-perceived latency) by input and output type. Dimensions