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.

This guide walks you through getting Mise running against your voice AI stack for the first time. You’ll connect your pipeline, ingest your first calls, and run a query that returns real matches clustered into defect signatures.
Mise is optimized for teams running 10,000+ calls per day. At that volume, acoustic indexing starts surfacing defect patterns that would be invisible in transcripts or call-level metrics.
1

Request access

Mise is in private alpha. Go to sf-voice.sh/sign-up and submit your team’s details — your stack, call volume, and the type of voice AI you’re running.The Mise team reviews requests by stack and volume. Once admitted, you’ll receive credentials and onboarding instructions by email.
Mention your primary stack (LiveKit, Twilio, Telnyx, Pipecat) and your daily call volume in the request form. This speeds up review.
2

Connect your voice stack

Mise integrates with the telephony and voice AI stacks your team already uses. Choose your stack and follow the integration guide:

LiveKit

Real-time voice and video pipelines

Twilio

Programmable telephony and voice

Telnyx

Cloud communications platform

Pipecat

Open-source voice AI orchestration
Each integration connects Mise to your call pipeline. Once connected, Mise begins ingesting calls: per-turn audio is archived alongside full call recordings, and acoustic indexing starts automatically.
Mise indexes audio at the turn level — each speaker turn is archived and processed independently. This is what makes turn-level replay and per-turn acoustic search possible.
3

Run your first query

Once calls are flowing, you can query your corpus using voice.query(). Queries are expressed in natural language and match on acoustic features, not just transcript keywords.
import mise

client = mise.Client(api_key="your-api-key")

# Search for calls where the agent interrupted the caller
results = client.voice.query("calls where the agent interrupted the caller")

print(f"{results.total_matches} matches")
print(f"{len(results.defect_signatures)} defect signatures")
You can also use the MCP server to query from Claude, Cursor, or Cline directly. See MCP Server.
Start with broad queries to understand what’s in your corpus, then narrow. Try:
  • "calls where the caller expressed frustration"
  • "calls where there was a long silence before the agent responded"
  • "calls where the caller repeated themselves more than twice"
4

Interpret your results

A voice.query() response contains two things: individual call matches and defect signatures.Matches are the specific calls that scored highest against your query. Each match includes:
  • The call record with metadata (from number, to number, timestamps, duration)
  • A sentiment score and summary
  • The specific turns that drove the match, with audio playback available
Defect signatures are automatically clustered groups of similar matches. Rather than reviewing 1,284 individual calls, Mise might cluster them into 6 distinct patterns — each with a name, a representative example, and a count.
for signature in results.defect_signatures:
    print(signature.name)          # e.g. "Agent interrupts during disambiguation"
    print(signature.count)         # e.g. 412
    print(signature.example_call)  # call ID for the most representative example
Use Call Replay to jump to the exact turn in a representative call and hear the audio alongside the transcript and tool call trace.
Defect signatures are clustered results from corpus search — Mise groups matches by acoustic similarity, not by keyword overlap. Two calls can match the same signature without sharing a single word.

​
What’s next

How it works

Understand the acoustic indexing model and the five dimensions Mise captures.

Corpus search

Learn the full query syntax and how to filter, rank, and export results.

Defect signatures

Understand how clustering works and how to act on signature output.

MCP server

Connect Claude, Cursor, or Cline to debug voice failures in natural language.