diff --git a/DESCRIPTION b/DESCRIPTION
index 318940f..3785530 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -1,5 +1,5 @@
 Package: compost
-Version: 0.1.0.12
+Version: 0.1.0.13
 Title: Video Compositing via 'FFmpeg'
 Description: Wraps common 'FFmpeg' <https://ffmpeg.org/> filter_complex
     patterns (overlay, chromakey, concat, scale, vstack) into clean R
diff --git a/NAMESPACE b/NAMESPACE
index 46d1981..07d2fd5 100644
--- a/NAMESPACE
+++ b/NAMESPACE
@@ -13,6 +13,7 @@ export(crossfade_concat)
 export(frame_export)
 export(frames_clip)
 export(hstack)
+export(normalize_audio)
 export(overlay)
 export(pad)
 export(pip)
diff --git a/R/audio.R b/R/audio.R
index 572a02c..8cb55e2 100644
--- a/R/audio.R
+++ b/R/audio.R
@@ -47,6 +47,59 @@ broadcast_audio <- function(input, output, dehum = TRUE, target_lufs = -14,
     invisible(output)
 }
 
+#' Normalize Audio Loudness
+#'
+#' A single loudnorm pass to an EBU R128 / ITU-R BS.1770 target -- the
+#' light-touch cousin of \code{\link{broadcast_audio}}, which wraps loudnorm
+#' in a full mastering chain (high-pass, de-hum notches, compression, fades).
+#' Use this to level takes that are already clean, e.g. TTS output across a
+#' batch of tracks.
+#'
+#' @param input Path to input audio (or A/V) file.
+#' @param output Path for output file (default: overwrite input in place).
+#' @param integrated Target integrated loudness in LUFS (default -16).
+#' @param lra Loudness range target in LU (default 11).
+#' @param tp True-peak ceiling in dBTP (default -1.5).
+#' @param bitrate Output audio bitrate for lossy formats (default "192k"),
+#'   or NULL for the encoder default.
+#' @param overwrite If TRUE (default), overwrite the output file.
+#' @param dry_run If TRUE, return the FFmpeg command without executing.
+#'
+#' @return Invisibly returns the output path. If dry_run, returns the
+#'   command string.
+#'
+#' @examples
+#' \dontrun{
+#' normalize_audio("audio.mp3")
+#' normalize_audio("take.wav", "leveled.wav", integrated = -14)
+#' }
+#'
+#' @export
+normalize_audio <- function(input, output = input, integrated = -16,
+                            lra = 11, tp = -1.5, bitrate = "192k",
+                            overwrite = TRUE, dry_run = FALSE) {
+    input <- normalizePath(input, mustWork = TRUE)
+    in_place <- identical(input, normalizePath(output, mustWork = FALSE))
+    out_path <- if (in_place && !dry_run) {
+        tempfile(fileext = paste0(".", tools::file_ext(input)))
+    } else {
+        normalizePath(output, mustWork = FALSE)
+    }
+
+    args <- c(if (overwrite) "-y", "-i", input, "-vn", "-af",
+              sprintf("loudnorm=I=%g:LRA=%g:TP=%g", integrated, lra, tp),
+        if (!is.null(bitrate)) c("-b:a", bitrate), out_path)
+
+    if (dry_run) {
+        return(.run_ffmpeg(args, dry_run = TRUE))
+    }
+    .run_ffmpeg(args)
+    if (in_place) {
+        file.rename(out_path, output)
+    }
+    invisible(output)
+}
+
 #' Build the -af filter graph for the broadcast-clean chain
 #'
 #' Pure string builder so the chain can be tested without ffmpeg.
diff --git a/inst/tinytest/test_audio.R b/inst/tinytest/test_audio.R
index 002aac3..77206f3 100644
--- a/inst/tinytest/test_audio.R
+++ b/inst/tinytest/test_audio.R
@@ -33,4 +33,27 @@ cmd <- broadcast_audio(wav, "clean.wav", dry_run = TRUE)
 expect_true(grepl("^ffmpeg ", cmd))
 expect_true(grepl("-af ", cmd))
 
+# normalize_audio: the loudnorm-only pass.
+cmdn <- normalize_audio(wav, "level.wav", dry_run = TRUE)
+expect_true(grepl("loudnorm=I=-16:LRA=11:TP=-1.5", cmdn))
+expect_true(grepl("-vn", cmdn))
+expect_true(grepl("-b:a 192k", cmdn))
+cmdn2 <- normalize_audio(wav, "level.wav", integrated = -14, bitrate = NULL,
+                         dry_run = TRUE)
+expect_true(grepl("loudnorm=I=-14:", cmdn2))
+expect_false(grepl("-b:a", cmdn2))
+
 unlink(wav)
+
+# In-place normalize on a real sine (at_home).
+if (at_home() && nzchar(Sys.which("ffmpeg"))) {
+    w <- tempfile(fileext = ".wav")
+    compost:::.run_ffmpeg(c("-y", "-f", "lavfi", "-i",
+                            "sine=frequency=440:duration=2", "-ar", "48000",
+                            "-ac", "1", w))
+    d0 <- probe(w, "duration")
+    normalize_audio(w)
+    expect_true(file.exists(w))
+    expect_true(abs(probe(w, "duration") - d0) < 0.1)
+    unlink(w)
+}
diff --git a/man/normalize_audio.Rd b/man/normalize_audio.Rd
new file mode 100644
index 0000000..d9bf60d
--- /dev/null
+++ b/man/normalize_audio.Rd
@@ -0,0 +1,44 @@
+% tinyrox says don't edit this manually, but it can't stop you!
+\name{normalize_audio}
+\alias{normalize_audio}
+\title{Normalize Audio Loudness}
+\usage{
+normalize_audio(input, output = input, integrated = -16, lra = 11, tp = -1.5,
+                bitrate = "192k", overwrite = TRUE, dry_run = FALSE)
+}
+\arguments{
+\item{input}{Path to input audio (or A/V) file.}
+
+\item{output}{Path for output file (default: overwrite input in place).}
+
+\item{integrated}{Target integrated loudness in LUFS (default -16).}
+
+\item{lra}{Loudness range target in LU (default 11).}
+
+\item{tp}{True-peak ceiling in dBTP (default -1.5).}
+
+\item{bitrate}{Output audio bitrate for lossy formats (default "192k"),
+or NULL for the encoder default.}
+
+\item{overwrite}{If TRUE (default), overwrite the output file.}
+
+\item{dry_run}{If TRUE, return the FFmpeg command without executing.}
+}
+\value{
+Invisibly returns the output path. If dry_run, returns the
+  command string.
+}
+\description{
+A single loudnorm pass to an EBU R128 / ITU-R BS.1770 target -- the
+light-touch cousin of \code{\link{broadcast_audio}}, which wraps loudnorm
+in a full mastering chain (high-pass, de-hum notches, compression, fades).
+Use this to level takes that are already clean, e.g. TTS output across a
+batch of tracks.
+}
+\examples{
+\dontrun{
+normalize_audio("audio.mp3")
+normalize_audio("take.wav", "leveled.wav", integrated = -14)
+}
+
+}