Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.sf-voice.sh/llms.txt

Use this file to discover all available pages before exploring further.

voice.query() accepts a plain English description of what you’re looking for across your call corpus. You don’t write filters or SQL — you describe a signal, and Mise maps it to the acoustic and semantic features indexed at every turn. 
voice.query("calls where the agent interrupted the caller.")
// → 1,284 matches
// → clustered into 6 defect signatures
Results include a total match count, a ranked list of individual calls, and automatically grouped defect signature clusters so you can prioritize the most common patterns first.

​
Basic syntax

Every query follows the same form: 
voice.query("<plain english description of what you're looking for>")
You can call it from your application code, the MCP server, or the Mise dashboard search bar. Queries are not case-sensitive and do not require special syntax. 
const results = await mise.voice.query(
  "calls where the caller sounded frustrated in the last two minutes"
);

​
Example queries

The table below shows how plain English maps to the acoustic dimensions Mise indexes.
QueryWhat it finds
"calls where the agent interrupted the caller"Rhythm: interruption events
"calls where the caller expressed frustration"Tension: frustration signal
"calls that ended without a resolution"Outcome: no_action or escalated_unresolved
"calls where the agent paused too long"Prosody: pause duration
"calls where the caller sounded uncertain"Tone: hesitancy
"calls where the caller asked to cancel"Intent: cancellation intent
"calls with high tension in the first 2 minutes"Tension: early escalation
"calls where both parties talked at the same time"Rhythm: overlap

​
Targeting acoustic features

You can query any of the five acoustic dimensions Mise indexes: tone, prosody, tension, rhythm, and intent. Name the signal in plain language — Mise resolves it to the indexed feature.
Tone queries target sentiment, irony, sarcasm, and emotional coloring.
voice.query("calls where the caller sounded angry")
voice.query("calls where the agent's tone was flat or dismissive")
voice.query("calls where the caller seemed satisfied at the end")
Prosody queries target pace, pauses, emphasis, and volume dynamics.
voice.query("calls where the agent spoke too quickly")
voice.query("calls where there were long silences after the caller asked a question")
voice.query("calls where the caller raised their voice")
Tension queries target frustration signals, escalation patterns, and de-escalation arcs.
voice.query("calls where tension escalated in the second half")
voice.query("calls where the agent successfully de-escalated the caller")
voice.query("calls where frustration was present throughout")
Rhythm queries target interruptions, overlap, turn-taking cadence, and conversation flow.
voice.query("calls where the agent frequently interrupted the caller")
voice.query("calls where both parties talked over each other")
voice.query("calls with unusually fast back-and-forth exchanges")
Intent queries target what the caller actually wanted, regardless of what they literally said.
voice.query("calls where the caller wanted to cancel but didn't say it directly")
voice.query("calls where the caller's goal was never identified")
voice.query("calls where the caller's intent shifted mid-conversation")

​
Targeting outcomes

Outcomes are resolved at the end of a call. You can query by what happened — or what didn’t. Outcome queries describe what happened at the end of the call. You can filter by resolution, escalation, cancellation, or lack of action. 
voice.query("calls that ended without a resolution")

voice.query("calls that were escalated and not resolved")

voice.query("calls where the caller cancelled")

voice.query("calls that were escalated to a human agent")
Outcome is determined by post-call summarization, not real-time detection. Queries targeting outcomes search completed calls only.

​
Combining conditions

You can combine multiple signals in a single query using natural language conjunctions. 
voice.query("calls where the agent interrupted the caller and the call ended unresolved")

voice.query("calls where the caller expressed frustration and asked to cancel")

voice.query("calls with high tension in the first two minutes and a long pause before escalation")
There is no limit on the number of conditions. Be as specific as you need — more precise descriptions return tighter, more actionable matches.

​
Scoping by time and duration

You can scope queries to a time window or a minimum call length. 
// Time window
voice.query("calls from last week where the agent paused too long")
voice.query("calls from the past 24 hours with unresolved outcomes")
voice.query("calls from Monday where tension escalated")

// Call duration
voice.query("calls longer than 5 minutes where the caller asked to cancel")
voice.query("short calls where the caller hung up quickly")
Timeframe scoping is especially useful when investigating a deployment change or a specific shift. Combine it with an acoustic or outcome condition to narrow results to a meaningful slice.

​
Scoping to a party

When the signal is specific to one side of the conversation, name the party explicitly. 
// Caller signals
voice.query("calls where the caller sounded frustrated")
voice.query("calls where the caller asked to speak to a human")

// Agent signals
voice.query("calls where the agent interrupted the caller")
voice.query("calls where the agent paused too long before responding")
Omitting the party returns results where either party matches the signal.

​
Results format

Every query returns three things:
1

Match count

The total number of calls in your corpus that match the query description.
2

Ranked calls

Individual calls ordered by relevance. Each result includes the call ID, timestamp, outcome, sentiment score, and a summary.
3

Defect signature clusters

Automatically grouped patterns across the matched calls. For example, 1,284 matches might cluster into six signatures such as “Agent interruption during complex requests” or “Caller frustration without resolution.” Use these clusters to prioritize fixes across your corpus rather than addressing individual calls.

​
Writing effective queries

Be specific about the signal

“Calls where the caller sounded frustrated” returns tighter results than “bad calls.” Name the acoustic dimension you care about.

Name the party

Specify “caller” or “agent” when the signal is one-sided. Unattributed queries match either party.

Include a timeframe

Add “from last week” or “in the past 24 hours” when investigating a specific incident or deployment window.

Use outcome language

Describe what happened at the end of the call — “ended without a resolution,” “was escalated,” “caller cancelled” — to filter by outcome.