replaygain: add ffmpeg backend

Add replaygain backend using ffmpeg's ebur128 filter.

The album gain is calculated as the mean of all BS.1770 gating block powers.
Besides differences in gating block offset, this should be equivalent to a
BS.1770 analysis of a proper concatenation of all tracks.

Just calculating the mean of all track gains (as implemented by the bs1770gain
backend) yields incorrect results as that would:
- completely ignore track lengths
  - just using length in seconds won't work either (e.g. BS.1770 ignores
    passages below a threshold)
- take the mean of track loudness, not power

When using the ffmpeg replaygain backend to create R128_*_GAIN tags, the
targetlevel will be set to -23 LUFS. GitHub PullRequest beetbox#3065 will make this
configurable.
It will also skip peak calculation, as there is no R128_*_PEAK tag.

It is checked if the libavfilter library supports replaygain calculation. Before
version 6.67.100 that did require the `--enable-libebur128` compile-time-option,
after that the ebur128 library is included in libavfilter itself. Thus we
require either a recent enough libavfilter version or the `--enable-libebur128`
option.

beetsplug/replaygain.py

@@ -18,6 +18,7 @@
 import subprocess
 import os
 import collections
+import math
 import sys
 import warnings
 import xml.parsers.expat
@@ -64,9 +65,22 @@ def call(args, **kwargs):
         raise ReplayGainError(u"argument encoding failed")
 
 
+def db_to_lufs(db):
+    """Convert db to LUFS.
+
+    According to https://wiki.hydrogenaud.io/index.php?title=
+      ReplayGain_2.0_specification#Reference_level
+    """
+    return db - 107
+
+
 # Backend base and plumbing classes.
 
+# gain: in LU to reference level
+# peak: part of full scale (FS is 1.0)
 Gain = collections.namedtuple("Gain", "gain peak")
+# album_gain: Gain object
+# track_gains: list of Gain objects
 AlbumGain = collections.namedtuple("AlbumGain", "album_gain track_gains")
 
 
@@ -81,11 +95,15 @@ def __init__(self, config, log):
         self._log = log
 
     def compute_track_gain(self, items):
+        """Computes the track gain of the given tracks, returns a list
+        of Gain objects.
+        """
         raise NotImplementedError()
 
     def compute_album_gain(self, items):
-        # TODO: implement album gain in terms of track gain of the
-        # individual tracks which can be used for any backend.
+        """Computes the album gain of the given album, returns an
+        AlbumGain object.
+        """
         raise NotImplementedError()
 
     def use_ebu_r128(self):
@@ -286,6 +304,257 @@ def use_ebu_r128(self):
         self.method = '--ebu'
 
 
+# ffmpeg backend
+class FfmpegBackend(Backend):
+    """A replaygain backend using ffmpegs ebur128 filter.
+    """
+    def __init__(self, config, log):
+        super(FfmpegBackend, self).__init__(config, log)
+        config.add({
+            "peak": "true"
+        })
+        self._peak_method = config["peak"].as_str()
+        self._target_level = db_to_lufs(config['targetlevel'].as_number())
+        self._ffmpeg_path = "ffmpeg"
+
+        # check that ffmpeg is installed
+        try:
+            ffmpeg_version_out = call([self._ffmpeg_path, "-version"])
+        except OSError:
+            raise FatalReplayGainError(
+                u"could not find ffmpeg at {0}".format(self._ffmpeg_path)
+            )
+        incompatible_ffmpeg = True
+        for line in ffmpeg_version_out.stdout.splitlines():
+            if line.startswith(b"configuration:"):
+                if b"--enable-libebur128" in line:
+                    incompatible_ffmpeg = False
+            if line.startswith(b"libavfilter"):
+                version = line.split(b" ", 1)[1].split(b"/", 1)[0].split(b".")
+                version = tuple(map(int, version))
+                if version >= (6, 67, 100):
+                    incompatible_ffmpeg = False
+        if incompatible_ffmpeg:
+            raise FatalReplayGainError(
+                u"Installed FFmpeg version does not support ReplayGain."
+                u"calculation. Either libavfilter version 6.67.100 or above or"
+                u"the --enable-libebur128 configuration option is required."
+            )
+
+        # check that peak_method is valid
+        valid_peak_method = ("true", "sample")
+        if self._peak_method not in valid_peak_method:
+            raise ui.UserError(
+                u"Selected ReplayGain peak method {0} is not supported. "
+                u"Please select one of: {1}".format(
+                    self._peak_method,
+                    u', '.join(valid_peak_method)
+                )
+            )
+
+    def compute_track_gain(self, items):
+        """Computes the track gain of the given tracks, returns a list
+        of Gain objects (the track gains).
+        """
+        gains = []
+        for item in items:
+            gains.append(
+                self._analyse_item(
+                    item,
+                    count_blocks=False,
+                )[0]  # take only the gain, discarding number of gating blocks
+            )
+        return gains
+
+    def compute_album_gain(self, items):
+        """Computes the album gain of the given album, returns an
+        AlbumGain object.
+        """
+        # analyse tracks
+        # list of track Gain objects
+        track_gains = []
+        # maximum peak
+        album_peak = 0
+        # sum of BS.1770 gating block powers
+        sum_powers = 0
+        # total number of BS.1770 gating blocks
+        n_blocks = 0
+
+        for item in items:
+            track_gain, track_n_blocks = self._analyse_item(item)
+
+            track_gains.append(track_gain)
+
+            # album peak is maximum track peak
+            album_peak = max(album_peak, track_gain.peak)
+
+            # prepare album_gain calculation
+            # total number of blocks is sum of track blocks
+            n_blocks += track_n_blocks
+
+            # convert `LU to target_level` -> LUFS
+            track_loudness = self._target_level - track_gain.gain
+            # This reverses ITU-R BS.1770-4 p. 6 equation (5) to convert
+            # from loudness to power. The result is the average gating
+            # block power.
+            track_power = 10**((track_loudness + 0.691) / 10)
+
+            # Weight that average power by the number of gating blocks to
+            # get the sum of all their powers. Add that to the sum of all
+            # block powers in this album.
+            sum_powers += track_power * track_n_blocks
+
+        # calculate album gain
+        if n_blocks > 0:
+            # compare ITU-R BS.1770-4 p. 6 equation (5)
+            # Album gain is the replaygain of the concatenation of all tracks.
+            album_gain = -0.691 + 10 * math.log10(sum_powers / n_blocks)
+        else:
+            album_gain = -70
+        # convert LUFS -> `LU to target_level`
+        album_gain = self._target_level - album_gain
+
+        self._log.debug(
+            u"{0}: gain {1} LU, peak {2}"
+            .format(items, album_gain, album_peak)
+            )
+
+        return AlbumGain(Gain(album_gain, album_peak), track_gains)
+
+    def _construct_cmd(self, item, peak_method):
+        """Construct the shell command to analyse items."""
+        return [
+            self._ffmpeg_path,
+            "-nostats",
+            "-hide_banner",
+            "-i",
+            item.path,
+            "-filter",
+            "ebur128=peak={0}".format(peak_method),
+            "-f",
+            "null",
+            "-",
+        ]
+
+    def _analyse_item(self, item, count_blocks=True):
+        """Analyse item. Returns a Pair (Gain object, number of gating
+        blocks above threshold).
+
+        If `count_blocks` is False, the number of gating blocks returned
+        will be 0.
+        """
+        # call ffmpeg
+        self._log.debug(u"analyzing {0}".format(item))
+        cmd = self._construct_cmd(item, self._peak_method)
+        self._log.debug(
+            u'executing {0}', u' '.join(map(displayable_path, cmd))
+        )
+        output = call(cmd).stderr.splitlines()
+
+        # parse output
+
+        if self._peak_method == "none":
+            peak = 0
+        else:
+            line_peak = self._find_line(
+                output,
+                "  {0} peak:".format(self._peak_method.capitalize()).encode(),
+                len(output) - 1, -1,
+                )
+            peak = self._parse_float(
+                output[self._find_line(
+                    output, b"    Peak:",
+                    line_peak,
+                )])
+            # convert TPFS -> part of FS
+            peak = 10**(peak / 20)
+
+        line_integrated_loudness = self._find_line(
+            output, b"  Integrated loudness:",
+            len(output) - 1, -1,
+            )
+        gain = self._parse_float(
+            output[self._find_line(
+                output, b"    I:",
+                line_integrated_loudness,
+            )])
+        # convert LUFS -> LU from target level
+        gain = self._target_level - gain
+
+        # count BS.1770 gating blocks
+        n_blocks = 0
+        if count_blocks:
+            gating_threshold = self._parse_float(
+                output[self._find_line(
+                    output, b"    Threshold:",
+                    line_integrated_loudness,
+                )])
+            for line in output:
+                if not line.startswith(b"[Parsed_ebur128"):
+                    continue
+                if line.endswith(b"Summary:"):
+                    continue
+                line = line.split(b"M:", 1)
+                if len(line) < 2:
+                    continue
+                if self._parse_float(b"M: " + line[1]) >= gating_threshold:
+                    n_blocks += 1
+            self._log.debug(
+                u"{0}: {1} blocks over {2} LUFS"
+                .format(item, n_blocks, gating_threshold)
+                )
+
+        self._log.debug(
+            u"{0}: gain {1} LU, peak {2}"
+            .format(item, gain, peak)
+            )
+
+        return Gain(gain, peak), n_blocks
+
+    def _find_line(self, output, search, start_index=0, step_size=1):
+        """Return index of line beginning with `search`.
+
+        Begins searching at index `start_index` in `output`.
+        """
+        end_index = len(output) if step_size > 0 else -1
+        for i in range(start_index, end_index, step_size):
+            if output[i].startswith(search):
+                return i
+        raise ReplayGainError(
+            u"ffmpeg output: missing {0} after line {1}"
+            .format(repr(search), start_index)
+            )
+
+    def _parse_float(self, line):
+        """Extract a float.
+
+        Extract a float from a key value pair in `line`.
+        """
+        # extract value
+        value = line.split(b":", 1)
+        if len(value) < 2:
+            raise ReplayGainError(
+                u"ffmpeg ouput: expected key value pair, found {0}"
+                .format(line)
+                )
+        value = value[1].lstrip()
+        # strip unit
+        value = value.split(b" ", 1)[0]
+        # cast value to as_type
+        try:
+            return float(value)
+        except ValueError:
+            raise ReplayGainError(
+                u"ffmpeg output: expected float value, found {1}"
+                .format(value)
+                )
+
+    def use_ebu_r128(self):
+        """Set this Backend up to use EBU R128."""
+        self._target_level = -23
+        self._peak_method = "none"  # R128 tags do not need peak
+
+
 # mpgain/aacgain CLI tool backend.
 class CommandBackend(Backend):
 
@@ -836,9 +1105,10 @@ class ReplayGainPlugin(BeetsPlugin):
         "gstreamer": GStreamerBackend,
         "audiotools": AudioToolsBackend,
         "bs1770gain": Bs1770gainBackend,
+        "ffmpeg": FfmpegBackend,
     }
 
-    r128_backend_names = ["bs1770gain"]
+    r128_backend_names = ["bs1770gain", "ffmpeg"]
 
     def __init__(self):
         super(ReplayGainPlugin, self).__init__()

docs/plugins/replaygain.rst

@@ -10,10 +10,10 @@ playback levels.
 Installation
 ------------
 
-This plugin can use one of three backends to compute the ReplayGain values:
-GStreamer, mp3gain (and its cousin, aacgain), Python Audio Tools. mp3gain
-can be easier to install but GStreamer and Audio Tools support more audio
-formats.
+This plugin can use one of many backends to compute the ReplayGain values:
+GStreamer, mp3gain (and its cousin, aacgain), Python Audio Tools or ffmpeg.
+mp3gain can be easier to install but GStreamer, Audio Tools and ffmpeg support
+more audio formats.
 
 Once installed, this plugin analyzes all files during the import process. This
 can be a slow process; to instead analyze after the fact, disable automatic
@@ -75,6 +75,15 @@ On OS X, most of the dependencies can be installed with `Homebrew`_::
 
 .. _Python Audio Tools: http://audiotools.sourceforge.net
 
+ffmpeg
+``````
+
+This backend uses ffmpeg to calculate EBU R128 gain values.
+To use it, install the `ffmpeg`_ command-line tool and select the
+``ffmpeg`` backend in your config file.
+
+.. _ffmpeg: https://ffmpeg.org
+
 Configuration
 -------------
 
@@ -106,6 +115,11 @@ These options only work with the "command" backend:
   would keep clipping from occurring.
   Default: ``yes``.
 
+This option only works with the "ffmpeg" backend:
+
+- **peak**: Either ``true`` (the default) or ``sample``. ``true`` is
+  more accurate but slower.
+
 Manual Analysis
 ---------------
 

test/test_replaygain.py

@@ -45,6 +45,8 @@
 else:
     LOUDNESS_PROG_AVAILABLE = False
 
+FFMPEG_AVAILABLE = has_program('ffmpeg', ['-version'])
+
 
 def reset_replaygain(item):
     item['rg_track_peak'] = None
@@ -205,6 +207,11 @@ def test_malformed_output(self, call_patch):
         self.assertEqual(len(matching), 2)
 
 
+@unittest.skipIf(not FFMPEG_AVAILABLE, u'ffmpeg cannot be found')
+class ReplayGainFfmpegTest(ReplayGainCliTestBase, unittest.TestCase):
+    backend = u'ffmpeg'
+
+
 def suite():
     return unittest.TestLoader().loadTestsFromName(__name__)