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.

A defect signature is a named, clustered pattern of similar failures across your call corpus. When you run a corpus search and get back 1,284 matches, Mise doesn’t leave you with a list of calls to review. It clusters those matches by acoustic and behavioral similarity into a small set of named signatures — typically 6 to 20 — each representing a distinct failure mode. Instead of asking “what went wrong in this call,” you ask “how many callers hit this pattern, and what’s the highest-impact one to fix first.”

​
How defect signatures are generated

Defect signatures are produced automatically from corpus search results.
1

Run a corpus search

Start with a natural language query that describes a class of failure. For example: voice.query("calls where the agent interrupted the caller").
2

Mise clusters the matches

Matching calls are grouped by acoustic similarity — tension arcs, rhythm patterns, turn-level sentiment progressions — not just textual similarity. Calls that sound alike cluster together even when the transcript content differs.
3

Signatures are named and ranked

Each cluster is named based on its dominant pattern and ranked by affected call count. You see the most impactful failures first.

​
What a defect signature contains

Each signature gives you enough context to understand, prioritize, and act on the pattern without reviewing individual calls.
A plain-language name and description of the failure pattern — what happens, when in the call it occurs, and which party is involved.Example: “Agent interruption during complex requests” — the agent cuts off the caller while the caller is mid-sentence on multi-step requests.
The number of calls in your corpus matching this signature. Used to prioritize which patterns to address first.
A curated set of calls that best illustrate the pattern. Each example links directly to the relevant turn in Call Replay.
The specific turns across affected calls where the pattern is most pronounced. Turn references include acoustic scores, transcript excerpts, and adjacent tool calls.

​
Example: interruption signatures

A query for "calls where the agent interrupted the caller" returns 1,284 matches clustered into 6 signatures:
SignatureCalls
Agent interruption during complex requests487
Agent interruption at handoff to human312
Agent interruption after long caller pause198
Double-speak overlap on greeting144
Agent interruption on repeat caller statement98
Agent interruption during cancellation request45
The call count tells you where to focus. The signature name tells you what’s happening. The representative examples let you verify before investing in a fix.

​
Using defect signatures

​
Prioritizing fixes by call count

Sort signatures by affected call count to identify the highest-impact patterns. A signature affecting 487 calls is a stronger candidate for immediate investigation than one affecting 45 — unless the 45-call pattern involves a critical outcome like unresolved escalations. 
const results = await voice.query("calls where the agent interrupted the caller");
const ranked = results.signatures.sort((a, b) => b.callCount - a.callCount);

ranked.forEach(sig => {
  console.log(`${sig.name}: ${sig.callCount} calls`);
});

​
Sharing with your AI assistant for root cause analysis

Defect signatures are structured, machine-readable output. You can pass a signature directly to Claude, Cursor, or Cline for root cause analysis. 
const sig = results.signatures[0];
const bundle = await sig.export();
// Returns: name, callCount, turnReferences, acousticSummary, representativeExamples
Paste a signature export into your AI assistant with a prompt like: “Here is a defect signature from our voice AI corpus. Analyze the turn references and acoustic scores, and suggest what’s causing this behavior.” The structured format means you don’t need to describe the data — the assistant reads it directly.

​
Tracking resolution over time

After you identify and address a defect, re-run the same corpus search query against new calls to measure whether the signature’s call count is declining. A shrinking count confirms the fix is working; a stable or growing count indicates the underlying behavior persists.
Corpus search always queries your full historical corpus. To measure improvement over time, filter by date range to compare call counts before and after a deployment.

​
Defect signatures and the MCP server

The Mise MCP server exposes defect signatures as resources your AI debugger can query and reason over directly. You don’t need to export data or format prompts — your AI assistant can inspect signatures, pull turn references, and request audio context inline.

MCP Server overview

Connect Claude, Cursor, or Cline to your call corpus using the Mise MCP server.

AI debugging guide

Use defect signatures and the MCP server to run structured root cause analysis.