Theory of Operation

Signal Chain at a Glance

Dashed border = optional stage. Dialog Guard, Level Riding, and Loudnorm run only when enabled in settings. The source file is never modified.

A Brief History of Film Sound

The audio embedded in a modern Blu-ray is the product of seventy years of competing standards, corporate rivalries, and engineering breakthroughs. Understanding the format tells you a great deal about why the audio sounds the way it does — and why it needs processing to work outside a cinema.

Why Film Audio Sounds Wrong at Home

Film audio is mixed in a room calibrated to a specific playback level. When you take that mix out of its intended environment, the dynamic range that was designed to create dramatic impact becomes an obstacle to simple comprehension.

The 85 dB SPL Reference

SMPTE RP200, the standard for theatrical sound calibration, specifies that pink noise at −20 dBFS should measure 85 dB SPL (C-weighted, slow) at the primary listening position — every channel, independently calibrated to the same level. This is the reference level at which theatrical mixers do their work.

With 20 dB of headroom above that reference, a theatrical system can produce 105 dB peak SPL per channel — louder than a rock concert, louder than a jet at 300 feet. The dynamic range between a whispered line of dialogue and a full-scale action sequence is typically 20–25 dB. By design.

Home Listening Reality

Home theater receivers are typically calibrated to 79–82 dB SPL at the listening position — 3 to 6 dB below theatrical reference. Headphone and laptop listening happens at 70–78 dB or lower. That gap isn't just about volume: the human ear's frequency and dynamic sensitivity changes with absolute playback level (the Fletcher-Munson effect), which means the perceived dynamic range shrinks as you turn down.

At 78 dB, a scene's dynamic range effectively compresses by about 7 dB. What read as a whisper and a shout in the theater becomes a murmur and a raised voice at home. The problem compounds with ambient noise — traffic, HVAC, household activity — masking the quiet end of the range entirely.

The Dialogue Problem

Film dialogue lives in the center channel, which in a 5.1 or 7.1 mix carries the primary voice frequencies (roughly 200 Hz – 4 kHz). Small speakers, earbuds, and laptop transducers have poor response in this midrange range. Music and effects tend to occupy lower and higher frequencies where small speakers often perform better — or at least differently — making dialogue feel recessed even when it's not.

The theatrical mix assumes you can hear at −31 LUFS because your room is quiet, your speakers are large, and you're calibrated. None of those assumptions hold on headphones during your commute.

The Numbers

A typical theatrical film integrates around −27 LUFS, which is Netflix's delivery standard. Spotify normalizes music to −14 LUFS. Apple Music targets −16 LUFS. That difference — up to 13 LUFS between a film mix and a streaming pop track — represents a factor of roughly 4.5× in average loudness. The film is simply much quieter by design, intended for a much louder room.

Typical Integrated Loudness Targets

Loudness normalization in FilmStrip bridges this gap: it brings the integrated loudness of the extracted audio up to a target that matches your listening context.

From Surround to Stereo

Extracting a 7.1 audio track from an MKV and rendering it as stereo requires folding eight channels into two. The math is defined by international standards and is not lossless in any perceptual sense — but it is principled.

The Channel Layout

A 7.1 mix contains eight discrete channels arranged around the listener:

The LoRo Downmix Matrix

FilmStrip uses a LoRo (Left-Only/Right-Only) downmix, the standard defined by ITU-R BS.775 and ATSC A/52. It is called "Left-Only/Right-Only" because the output channels are independently derived — there is no phase encoding between them. The result plays correctly on any stereo system without special decoding.

The alternative is LtRt (Left Total/Right Total), which uses phase manipulation to encode surround information in a way that a Dolby Pro Logic decoder can later extract. LtRt produces a wider apparent soundstage on a compatible system but degrades on simple stereo playback. For a file you intend to listen to on headphones or a stereo player, LoRo is correct.

Implementation note: rather than relying on FFmpeg's default downmix matrix, FilmStrip applies an explicit pan filter so the center-channel weight can be set deliberately. The coefficients are:

Why These Gain Values?

The textbook LoRo matrix mixes the center channel at 0.707 (−3 dB) — the equal-power sum point. The argument: if FC were summed at full gain alongside content in FL, the sum at the left output could exceed either source alone by 3 dB, risking a louder and spatially confused image. Mixing FC at 0.707 keeps the front soundstage balanced.

FilmStrip deviates from this convention and folds FC at unity gain. Films, unlike music, place dialog almost exclusively in the center channel (see Dialog Guard, below), and a 3 dB attenuation makes voices feel recessed in stereo playback. The brick-wall limiter at the end of the chain catches any inter-sample peaks that result from summing FC at full level.

The side channels get an additional −3 dB reduction (0.5, or −6 dB total) because they carry primarily ambient content — room tone, diffuse reverb, passing sounds. Folding them in at full surround level would fill the stereo image with diffuse content and reduce its clarity.

What You Actually Hear

After the downmix, dialogue arrives equally in both channels at unity gain, placing it as a centered phantom image between the speakers and at the same level it had in the original mix. Music and scored elements retain their left/right panning from the front channels. Surround ambience folds into both channels at reduced level, adding a sense of space without dominating the image. Sub-bass effects disappear — explosions and impacts remain audible through the full-range channels but lose their low-frequency weight.

The result is a serviceable, intelligible stereo mix that prioritizes the most important element — dialogue — and preserves the music and primary effects in a form that translates well to headphones.

The Processing Pipeline

Each step runs in sequence, writing intermediate results to a temporary directory. The source file is never modified.

Pass 1: Extraction

ffmpeg decodes the selected audio stream. If Dialog Guard is enabled on a 5.1 or 7.1 source, it first splits out the center channel, normalizes it independently, and reassembles the multichannel signal before the downmix. For stereo or mono sources, Stereo Dialog Assist performs the equivalent operation on the mid (L+R) signal. If the High Pass Filter is enabled, two cascaded 80 Hz biquads (highpass=f=80,highpass=f=80) run next, rolling off subsonic content at 24 dB/oct before the downmix. The filter chain then downmixes to stereo via an explicit pan filter using the matrix described in From Surround to Stereo, resamples to 44.1 kHz, and applies Level Riding (dynaudnorm) on the stereo mix when enabled. A brick-wall limiter (alimiter=limit=0.99:attack=5:release=50:level=false) runs last to catch any inter-sample peaks from channel summation. The result is a temporary 24-bit PCM WAV file.

24-bit depth is chosen rather than 16-bit to preserve headroom for subsequent processing steps. Loudness normalization may apply a gain of several dB; 24 bits provides 144 dB of dynamic range so there is no meaningful resolution loss at any gain setting.

Pass 2 (optional): Loudness Analysis

The loudness normalization process requires two ffmpeg passes. The first pass runs ffmpeg with the loudnorm filter in analysis mode, outputting its measurements to /dev/null — no audio is written. At the end of the pass, the filter prints a JSON block to stderr containing the measured integrated loudness (LUFS), true peak (dBTP), loudness range (LRA), and threshold values.

This analysis runs against the already-downmixed and (optionally) level-ridden WAV from Pass 1 — so the loudness measurement reflects the final dynamics of the audio, not the original theatrical mix.

Pass 3 (optional): Loudness Normalize

The second pass provides the measured values back to the loudnorm filter along with the target LUFS. With linear=true, the filter computes a single static gain value: the difference between the measured integrated loudness and the target, accounting for true peak ceiling. That gain is applied uniformly across the entire file.

Linear mode preserves dynamics exactly. The output has the same waveform shape as the input — only louder or quieter by a fixed amount. This is ideal after level riding, which has already done the dynamic work.

Level Riding: Dynamic Audio Normalizer

Before automated systems existed, a broadcast engineer sat at a console and manually rode the fader — turning down peaks, pulling up quiet passages — to keep audio levels consistent for home listeners. The goal was the same as it is today: tame the dynamic range of content mixed for a different context. dynaudnorm is a modern, principled version of that process.

Frame Analysis

The filter divides the audio into short frames — 500 milliseconds by default. For each frame, it computes the peak magnitude. This is not RMS (average energy) but true peak — the highest instantaneous sample level in the frame. The gain needed to bring that peak to the target level p is calculated: gain = p / peak.

If the result exceeds the maximum allowed gain m, the gain is clamped to m. This prevents very quiet frames (near-silence, room noise) from being boosted so aggressively that the noise floor becomes intrusive.

Gaussian Smoothing

A raw per-frame gain would cause audible pumping — volume jumping every half-second in sync with the analysis window. dynaudnorm prevents this by computing a Gaussian-weighted average across a neighborhood of frames (the window size g) and using the smoothed gain value instead of the raw one.

The Gaussian weighting means frames near the current position in time have more influence than frames far away. With the default window of 31 frames (about 15.5 seconds of lookahead and lookbehind), gain transitions are spread over several seconds — imperceptible as discrete steps. The filter effectively "knows" what's coming and begins adjusting early.

The p Parameter: Peak Target

The peak target controls how aggressively loud frames are attenuated. A value of p = 0.95 means the loudest frame in each window will be brought to 95% of full scale — very gentle, leaving most of the original dynamic range intact. At p = 0.55, loud frames are pulled down to 55% of full scale — a reduction of about −5.2 dB — noticeably more aggressive.

The m Parameter: Maximum Gain

This is the parameter that actually closes dynamic range rather than just attenuating peaks. m sets the maximum gain that can be applied to a frame. At m = 1.0, gain can never exceed 1.0 — the filter can only attenuate loud frames, never boost quiet ones. Quiet dialogue stays quiet.

At m = 10.0, the filter can amplify a quiet frame by up to 10× (roughly +20 dB). Scenes where a character whispers, or where the only content is ambient room tone, get boosted significantly — narrowing the perceived gap between quiet and loud. This is the effective dynamic range compression that makes film audio comfortable to listen to without a home theater.

Headphone-Tuned Settings

FilmStrip ships a single, fixed tuning chosen for headphone listening:

This tuning is effectively downward-only. With m = 1.5, the filter still passes the dynamics of a typical scene through largely intact: dialogue at −20 dBFS needs a gain of 4.5× to reach the target, so it stays where the mix put it. The few decibels of available upward gain rescue brief whispers and quiet exchanges without amplifying ambient hiss or HVAC rumble during a silent intro. The wider work of bringing quiet film dialog forward in the mix is handled separately by Dialog Guard (5.1/7.1) and Stereo Dialog Assist (stereo/mono), which act on the center / mid signal where dialog actually lives.

Dialog Guard: Center Channel Normalization

Film dialog lives almost exclusively in the center channel — a deliberate mixing decision that dates to the earliest days of surround sound. When a 5.1 or 7.1 mix is downmixed to stereo, the center channel is attenuated by −3 dB before being summed into both output channels. If a scene's dialog was already mixed quietly relative to the action, that attenuation can push it below the threshold of comfortable comprehension. Dialog Guard addresses this at the source, before the downmix happens.

Why the Center Channel Specifically

In any standard 5.1 or 7.1 theatrical mix, the center channel (FC) carries dialog with essentially no other content. Explosions are in the LFE. Music is spread across FL and FR. Ambience and effects fill BL, BR, SL, and SR. The center is reserved for voices — and in the channel layout for both 5.1 and 7.1, FC is always index 2, making it unambiguously targetable regardless of the specific surround configuration.

Full-mix normalization tools like dynaudnorm with a wide window can miss brief quiet passages — a character whispering for a few seconds in an otherwise loud scene, for instance. The window average is dominated by the loud content surrounding the passage, so the gain assigned to that window remains low. The whisper stays quiet. Targeting the center channel independently, with a shorter window, allows these brief dips to be caught and corrected before they get buried further by the downmix.

How It Works

When Dialog Guard is enabled on a 5.1 or 7.1 source, the entire extraction pass switches from a simple -af filter chain to a -filter_complex graph that processes channels as parallel streams:

The reassembled multichannel signal then continues through the normal filter chain: optional level riding, resample to 44.1 kHz, stereo downmix, and peak limiting. The center channel enters the downmix already normalized — the −3 dB attenuation in the downmix matrix is applied to a signal that has already been leveled, rather than to one that may have been quietly mixed to begin with.

Window Size: g=15

The Gaussian window of 15 frames covers approximately 0.5 seconds — half the default 31-frame window used by Level Riding. This shorter window allows the filter to react to brief quiet passages that a wider window would smooth over. The tradeoff is slightly more audible gain transitions, but on a mono channel carrying dialog the effect is rarely perceptible.

Maximum Gain: m=3×

The maximum gain of 3× (~+9.5 dB) is applied only when the center channel's peak is very low — near-silence or extremely quiet dialog. In practice, the filter boosts passages by a few dB in most cases. The ceiling is set low enough that ambient room tone or background noise on the center channel doesn't become audible hiss during pauses in dialog, but high enough to lift a whispered line into intelligibility.

Stereo Dialog Assist

Stereo and mono sources have no separable center channel, but dialog still concentrates in the centered (mono) component of the mix. Stereo Dialog Assist performs an equivalent operation by decomposing the stereo signal into mid (sum) and side (difference) components, applying the same dynaudnorm=p=0.88:m=3:g=15 filter to the mid signal only, and reassembling. The side signal — which carries the stereo width: ambience, music spread, panned effects — is untouched, so the stereo image stays intact while the dialog (which sits in the mid) gets normalized. Mono sources are treated as pure-mid and run the same filter across the entire signal.

Works With Level Riding

Using Dialog Guard (or Stereo Dialog Assist) alongside Level Riding is the intended combination. The dialog-side filter normalizes a single mono signal where dialog lives, at a fast window optimized for short passages. Level Riding then normalizes the stereo mix after the downmix, at a slower window that smooths transitions across longer scenes. They operate at different stages, on different signals, for different purposes.

Loudness Normalization: EBU R128

For most of the digital audio era, engineers chased loudness by bricking signals against 0 dBFS with limiters. The result was audio that was measurably distorted, perceptually fatiguing, and impossible to compare across sources. EBU R128 changed the unit of measurement from peak level to perceived loudness — and eventually ended the war.

The Loudness War

The problem began with CD mastering in the 1990s. Peak normalization ensured that no sample exceeded 0 dBFS, and for a brief period, that was sufficient. Then labels began competing for shelf presence: a song that sounded louder in a record store — even by a few dB — was perceived as more powerful, more professional, more worth buying.

Engineers responded by compressing dynamic range and limiting peaks with increasing aggression, allowing average levels to rise without exceeding 0 dBFS. By 2008, some releases averaged −9 dBFS — saturated, clipping, often visibly distorted in a waveform editor. Metallica's Death Magnetic (2008) became the canonical example of extreme loudness war damage: the album version famously measured worse on dynamic range tests than the Guitar Hero video game version of the same recording.

The problem was the measurement. Peak level tells you nothing about how loud something sounds to a human listener. Two signals can share the same peak level and differ by 10 dB in perceived loudness. A new measurement standard was needed.

ITU-R BS.1770 and K-Weighting

The International Telecommunication Union published BS.1770 as a method for measuring the subjective loudness of audio programs. The core of the algorithm is K-weighting — a frequency filter that approximates how the human auditory system responds to sound at normal listening levels.

K-weighting applies two stages of filtering before the RMS measurement:

After K-weighting, the algorithm measures mean square energy across overlapping 400 ms gating blocks, applies a two-stage threshold to exclude silence and near-silence, and reports the result in LUFS — Loudness Units relative to Full Scale. LUFS and LKFS (the ITU acronym: Loudness, K-weighted, relative to Full Scale) are identical measurements; the naming convention differs between the ITU and EBU standards but the values are interchangeable.

EBU R128: Loudness by Law

The European Broadcasting Union's Recommendation R128, first published in 2010, applied the ITU-R BS.1770 measurement algorithm to broadcast and mandated a target of −23 LUFS integrated loudness for all content delivered to European broadcasters. For the first time, a regulatory body had replaced peak normalization with loudness normalization as the delivery standard.

The immediate effect was that a heavily compressed pop record and a lightly processed spoken-word program, if both mastered to −23 LUFS, would play at the same perceived loudness on broadcast. The incentive to compress was removed: you could not get louder on air than a competitor by bricking your track harder, because the broadcaster's normalizer would bring both to the same level.

Streaming platforms followed. Spotify, Apple Music, YouTube, and others all adopted loudness normalization with targets ranging from −14 to −23 LUFS depending on their playback philosophy. The loudness war, at least for the formats that carry normalization metadata, was over.

Two-Pass Normalization

Accurate loudness normalization requires knowing the measured loudness before applying the correction — you cannot normalize in a single real-time pass without reading the entire file first. FilmStrip uses ffmpeg's loudnorm filter in two-pass mode.

Pass 1 — Analysis: ffmpeg processes the full audio file with the loudnorm filter in measurement mode, writing output to /dev/null. At the end, the filter reports the measured integrated loudness (I), true peak (TP), loudness range (LRA), and gating threshold. For a two-hour film this pass takes time proportional to the duration.

Pass 2 — Normalize: The second pass provides the measured values back to the loudnorm filter along with the target LUFS. With linear=true, the filter computes a single static gain value. That gain is applied uniformly across the entire file. Linear mode preserves dynamics exactly — the output has the same waveform shape as the input, only louder or quieter by a fixed amount.

Choosing a Target

The LUFS target determines how loud the output sits relative to other audio on your device. The right value depends on your listening context:

If the measured loudness is already below the true-peak ceiling (−1 dBTP by default), the normalization is a simple gain increase. If the source peaks would clip at the target loudness, the gain is reduced slightly to keep the output within the ceiling — so the integrated loudness may land slightly below target on very dynamic material. This is correct behavior: it prevents digital clipping at the cost of fractionally missing the loudness target.

Settings Reference

Output Format

Dialog Guard

Stereo Dialog Assist

High Pass Filter

Level Riding

Loudness Normalization

Choosing a Source Codec

Not all audio codecs are equal inputs for FilmStrip. When you have a choice of which track to use, the codec embedded in the source determines whether processing introduces an extra lossy transcode or not. The four you'll encounter most often on Blu-ray rips are AAC, E-AC3, AC3, and DTS.

AAC — Best Choice

AAC is the same codec FilmStrip produces for M4A output. If you're exporting to M4A, selecting an AAC source track means no transcode at all for that format — the file stays within the AAC family the entire time. WAV output decodes and re-encodes regardless of source, so the advantage is most significant for M4A output. When an AAC track is available, always prefer it.

E-AC3 — Second Choice

Dolby Digital Plus (E-AC3) is the dominant format on modern Blu-ray rips. It supports up to 1.5 Mbps and carries full 7.1 surround. Quality is high at typical encode bitrates. The decode → process → re-encode chain introduces a single generation of lossy conversion, which is audibly transparent at 192+ kbps M4A output. E-AC3 at 640 kbps is an excellent source.

AC3 — Acceptable

Dolby Digital (AC3) is the older 5.1 format with a ceiling of 640 kbps. It's perceptually lossier than E-AC3 at comparable bitrates, but in practice most AC3 tracks at 448–640 kbps sound fine after extraction. The extra generation loss is the same as E-AC3 — one decode and one re-encode. If AC3 is the only option, it will work well.

DTS — No Practical Advantage

DTS-HD Master Audio carries lossless audio on Blu-ray, but MKV rips typically demux to lossy DTS Core at 1.5 Mbps. The decode → process → re-encode chain is identical to E-AC3 — FilmStrip produces the same output quality regardless. There is no reason to seek out DTS specifically; E-AC3 or AAC are better or equal in every case.

Priority Order

• 1
  AAC — no transcode for M4A output; always prefer when available.
• 2
  E-AC3 — high bitrate ceiling, widely available on modern rips, one generation loss.
• 3
  AC3 — older format, lower bitrate ceiling, same number of generations as E-AC3.
• 4
  DTS — no quality advantage over E-AC3 in practice; treat it the same.