diff --git a/src/docgen/align.py b/src/docgen/align.py
index 48be636..79b434c 100644
--- a/src/docgen/align.py
+++ b/src/docgen/align.py
@@ -129,6 +129,19 @@ def probe_duration(audio_path: Path) -> float:
     return _finite_probe_seconds(audio_path, out.stdout.strip())
 
 
+def _require_silence_settings(noise_db: float, min_silence_sec: float) -> None:
+    """Silence settings must be finite, and the minimum gap must be positive."""
+    if not math.isfinite(noise_db):
+        raise AlignmentError(
+            f"silence noise must be a finite number of dB, not {noise_db!r}"
+        )
+    if not math.isfinite(min_silence_sec) or min_silence_sec <= 0:
+        raise AlignmentError(
+            "min silence must be a finite number of seconds greater than zero, "
+            f"not {min_silence_sec!r}"
+        )
+
+
 def detect_speech_intervals(
     audio_path: Path,
     duration: float,
@@ -137,6 +150,7 @@ def detect_speech_intervals(
     min_silence_sec: float = DEFAULT_MIN_SILENCE_SEC,
 ) -> list[tuple[float, float]]:
     """Run ffmpeg silencedetect and return speech (non-silent) intervals."""
+    _require_silence_settings(noise_db, min_silence_sec)
     cmd = [
         "ffmpeg", "-hide_banner", "-nostats",
         "-i", str(audio_path),
diff --git a/tests/test_align.py b/tests/test_align.py
index 648c7f9..ef48d4f 100644
--- a/tests/test_align.py
+++ b/tests/test_align.py
@@ -5,7 +5,9 @@
 import pytest
 
 from docgen.align import (
+    AlignmentError,
     build_local_timing,
+    detect_speech_intervals,
     parse_silencedetect_output,
     reconcile_intervals_to_sentences,
     split_sentences,
@@ -191,3 +193,17 @@ class _Proc:
 
     monkeypatch.setattr("docgen.align.subprocess.run", lambda *_a, **_k: _Proc())
     assert detect_speech_intervals(Path("/tmp/x.mp3"), 10.0) == [(0.0, 10.0)]
+
+
+def test_silence_settings_must_be_finite(monkeypatch) -> None:
+    from pathlib import Path
+
+    def _boom(*_a, **_k):
+        raise AssertionError("ffmpeg should not run")
+
+    monkeypatch.setattr("docgen.align.subprocess.run", _boom)
+    audio = Path("/tmp/x.mp3")
+    with pytest.raises(AlignmentError, match="min silence"):
+        detect_speech_intervals(audio, 10.0, min_silence_sec=float("inf"))
+    with pytest.raises(AlignmentError, match="silence noise"):
+        detect_speech_intervals(audio, 10.0, noise_db=float("nan"))