A GitHub Action That Fails the Build When Your Voice Agent Gets Worse

In the last post we put call audio in the test suite: treat a few golden
recordings as fixtures, analyze them on every change, and assert on the report so
a regression turns the build red. That works — but it's a pile of pytest you
have to write and maintain.

This post skips the boilerplate. There's now a GitHub Action that does the whole
thing: AudioTrace regression gate.
Point it at a folder of recordings and a committed baseline, and it fails the
build when call quality regresses — latency, sentiment, drop-off, cost, or
compliance. It runs entirely on open models, so there are no secrets and no
network calls in CI.

Here's the whole thing, start to finish.

The idea in one sentence

Voice agents drift in ways no unit test catches — a prompt tweak makes the agent
slower, a new TTS voice makes it colder, a refactor drops a required disclosure.
So we commit a baseline of what "good" sounds like, and gate every PR
against it.

Step 1 — Commit a baseline

Grab a handful of representative call recordings — a happy path, a frustrated
caller, a compliance-heavy call — and drop them in your repo (say tests/calls/).
Then generate a baseline locally:

pip install audiotrace          # FFmpeg must be on your system
audiotrace baseline tests/calls -o baseline.json

This analyzes every recording and writes baseline.json — the committed snapshot
of "good." Commit it alongside your code:

git add tests/calls baseline.json
git commit -m "Add voice-quality baseline"

That's your golden master. When you intentionally improve the agent, you
regenerate and re-commit it — same workflow as snapshot testing.

Step 2 — Add the Action

Drop this into .github/workflows/voice-quality.yml:

name: Voice quality
on: [pull_request]

jobs:
  audiotrace:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: dimastatz/audiotrace@v1
        with:
          calls: tests/calls
          baseline: baseline.json

Two required inputs — where the recordings live and where the baseline lives.
That's it. The Action installs AudioTrace + FFmpeg, re-analyzes the calls,
compares each against the baseline, and exits non-zero on any out-of-tolerance
regression.

Step 3 — Watch it catch a regression

Open a PR that changes a prompt or swaps the model. On the next run the gate
re-scores your golden calls and, if the agent got measurably worse, the check
fails with a summary like:

FAIL  frustrated_customer.wav
  Response latency (p95)  1,820ms → 2,540ms   (+39%, allowed +15%)
  Sentiment               0.12 → -0.20        (Δ0.32, allowed 0.10)

1 of 3 calls regressed.

A green build means the change was safe to ship. A red build means you made the
agent slower or colder before a customer felt it.

Every run also uploads a per-call HTML + JSON report as a build artifact —
even on failure — so you can open the report and see exactly what moved.

Inputs

Input	Required	Default	What it does
calls	yes	—	Directory of golden call recordings to gate.
baseline	yes	baseline.json	The committed baseline to compare against.
report-dir	no	audiotrace-report	Where per-call HTML + JSON reports are written.
version	no	audiotrace	pip spec to pin AudioTrace (e.g. audiotrace==1.2.2).
python-version	no	3.12	Python version the gate runs on.

Tuning the tolerances

Conversational signals wiggle run to run, so the gate ships with sane per-metric
tolerances — a band the number can move within before it counts as a regression:

Metric	Tolerance
Quality score	±0.05
Sentiment	±0.10
Response p95	+15%
Cost	+20%
Interruptions	+1
Frustration / drop-off / compliance	zero — any regression fails

New recordings that aren't in the baseline yet are skipped, not failed, so
adding fixtures never breaks the build.

Why this runs in CI at all

The trick that makes this practical: audiotrace analyze() runs locally on open
models (Whisper, pyannote, Librosa) — no hosted API, no key, no per-call bill. So
the gate needs nothing but your recordings and a runner. That's what lets it live
in pull_request CI instead of a nightly job behind a secret.

Try it

• Action: AudioTrace regression gate on the Marketplace
• Library: pip install audiotrace · github.com/dimastatz/audiotrace

Point it at three recordings and a baseline, open a PR, and watch a bad prompt
change go red. That's regression testing for the part of your product that used
to be a black box.