Call metrics

Every completed call computes a set of metrics from its transcript and recording. This page lists each one by its field name, what it measures, and which channel (voice, chat, or both) it applies to, plus the transcript’s speaker roles. For how these numbers roll up across a whole run, see Analytics & metrics; for reading one call’s transcript in the UI, see Calls & transcripts.

Overall score

Field	Channel	Meaning
overall_score	Voice	A CSAT (customer satisfaction) score from 1 to 10, evaluated from the call recording. If that evaluation can’t be parsed, the call falls back to a pass/fail signal reported by the voice provider instead
overall_score	Chat	The same field, but computed as a CSAT score from the transcript directly, since there’s no recording to evaluate

CSAT is the label you’ll see on this field wherever it’s surfaced in the product.

Talk time and interruptions

Voice-only. Computed from how much of the call each party spent speaking.

Field	Unit	Meaning
talk_ratio	ratio	How much of the call the agent spent talking versus the caller. Shown as a split between agent talk percentage and customer talk percentage
user_interruption_count	count	How many times the caller interrupted the agent
user_interruption_rate	rate	How often the caller interrupted, as interruptions per call
ai_interruption_count	count	How many times the agent interrupted the caller
ai_interruption_rate	rate	How often the agent interrupted, as interruptions per call
avg_stop_time_after_interruption_ms	ms	Average time the interrupted party takes to stop talking once interrupted

Speaking pace

Voice-only.

Field	Unit	Meaning
user_wpm	words/min	The caller’s speaking pace
bot_wpm	words/min	The agent’s speaking pace

Latency and response time

Field	Channel	Unit	Meaning
avg_agent_latency_ms	Voice	ms	Average time the agent takes to respond after the caller stops talking
avg_latency_ms	Chat	ms	Average time the agent takes to respond after the user’s message
response_time_ms	Voice	ms	Average duration of the agent’s own turns, computed from the transcript. This measures how long the agent’s replies run, not how quickly it starts them

Duration

Field	Unit	Meaning
duration_seconds	seconds	Length of the call. Voice calls are capped at 1800 seconds (30 minutes)

A run’s total duration is the sum of its calls’ duration_seconds, not a separately measured value.

Turn count and token usage

Chat-only.

Field	Unit	Meaning
turn_count	count	Number of back-and-forth exchanges in the conversation
total_tokens	tokens	Total tokens consumed by the agent’s language model calls during the chat
input_tokens	tokens	Tokens sent to the model as input
output_tokens	tokens	Tokens generated by the model as output

Cost breakdown

Field	Meaning
cost_cents	Total cost of the call, in cents
stt_cost_cents	Cost of converting the caller’s speech to text
llm_cost_cents	Cost of the language model calls that drove the conversation
tts_cost_cents	Cost of converting the agent’s replies to speech
storage_cost_cents	Cost of storing the call recording

Transcript and speaker roles

Each turn in a call’s transcript carries its text content, a start and end timestamp in milliseconds, and, for voice calls, a confidence score from speech recognition.

Speaker role	Shown in the transcript view	Meaning
USER	Yes	The caller or chat user’s turn
ASSISTANT	Yes	The agent’s turn
SYSTEM	Yes	A system-level turn in the conversation
TOOL_CALLS	No	Records which tool the agent invoked. Feeds tool-call evaluation rather than the visible transcript
TOOL_CALL_RESULT	No	The result a tool call returned. Also feeds tool-call evaluation, not the visible transcript
UNKNOWN	No	A turn that didn’t match any of the above; rare