CineAudioDB

A real-world evaluation set for cinematic audio source separation.

CineAudioDB contains real film/animation productions with ground-truth stems (dialogue, music, sfx) for evaluating cinematic source-separation models. Unlike synthetic datasets that sum stems linearly, real productions are mixed with a non-linear mastering chain (compression, limiting, sidechain ducking, reverb), so the released stems do not generally sum to the mastered mix. To support fair evaluation under both assumptions, every production is provided in two input versions:

• linear_mix — the three stems summed (perfectly additive: mix = dialogue + music + sfx).
• production_mix — the real mastered stereo mix (non-additive; this is what audiences hear).

This is the real-content evaluation companion to the synthetic generation pipeline described in the accompanying paper.

At a glance

Productions	9 (7 open-source animations + 2 student production)
Stems	dialogue, music, sfx (ambience merged into sfx); film_01 also ships 8 fine stems → 3-/4-stem groupings
Mix versions	linear_mix (additive), production_mix (mastered, where available)
Total duration	2.46 h
Audio format	FLAC (lossless), 24-bit, 48 kHz
Channels	stereo (one production is mono native)

Why two mix versions

Mask-based separators (e.g. Bandit, MRX) assume mix = Σ stemₖ and estimate each source as ŝₖ = x ⊙ mₖ. Real mastered mixes break this assumption: mastering redistributes and limits energy so individual stems routinely exceed the mix in some time-frequency bins.

• Evaluate on linear_mix to isolate model quality from the mixing-process mismatch (the input decomposes exactly into the ground-truth stems).
• Evaluate on production_mix to measure real deployment performance on professionally mastered audio. Because the mapping is non-additive, scale-/shift-sensitive metrics (e.g. SI-SNR) are best complemented by listening tests.

Directory structure

CineAudioDB/
├── README.md
├── NOTICE.md            # source attributions (CC-BY)
├── metadata.csv         # one row per audio file (HF audiofolder-compatible)
├── films.json           # per-production summary
└── data/
    ├── film_01/         # dialogue, music, sfx, linear_mix          (additive)
    │   └── fine/        # 8 atomic print-master stems (see below)
    ├── film_02/         # dialogue, music, sfx, linear_mix, production_mix
    ├── film_03/         # dialogue, music, sfx, linear_mix, production_mix   (mono)
    └── film_04/         # dialogue, music, sfx, linear_mix, production_mix

Within a production, every file has identical length, sample rate, and channel count.

Fine stems (film_01)

film_01 is delivered as a full professional print master, so it uniquely ships its 8 atomic stem categories in data/film_01/fine/ — dx (dialogue), vo (voice-over), mx (music), fx (hard effects), fol (foley), bg (backgrounds/ambience), pfx (production effects), arch (archival). They are scaled by the same per-production gain as the 3-stem files, so any grouping sums exactly (to 24-bit precision) to the corresponding stem and to linear_mix:

grouping	dialogue	music	ambience	sfx
3-stem (top-level files)	dx+vo	mx	—	fx+fol+bg+pfx+arch
4-stem (paper's synthesis taxonomy)	dx+vo	mx	bg	fx+fol+pfx+arch

import soundfile as sf, numpy as np
fine = "data/film_01/fine/"
def s(*names): return sum(sf.read(fine+n+".flac")[0] for n in names)
dialogue, music, ambience, sfx = s("dx","vo"), s("mx"), s("bg"), s("fx","fol","pfx","arch")
# dialogue + music + ambience + sfx == data/film_01/linear_mix.flac  (24-bit precision)

Productions (anonymized)

id	duration	channels	mix type	production_mix	source	license
film_01	78.00 min	stereo (downmix from 5.1)	additive¹	—	student production	CC-BY-SA-4.0
film_02	12.11 min	stereo	non-additive	✓	open-source animation	CC-BY-4.0
film_03	10.45 min	mono	non-additive	✓	open-source animation	CC-BY-4.0
film_04	3.22 min	stereo (DAW reconstruction)	non-additive	✓	open-source animation	CC-BY-3.0
film_05	7.58 min	stereo (Blender VSE render)	non-additive	✓	open-source animation (EN)	CC-BY-SA-4.0
film_06	7.59 min	stereo (Blender VSE render)	non-additive	✓	open-source animation (RU)	CC-BY-SA-4.0
film_07	1.89 min	stereo (Blender VSE render)	additive³	—	open-source animation (EN)	CC-BY-SA-4.0
film_08	1.90 min	stereo (Blender VSE render)	additive³	—	open-source animation (RU)	CC-BY-SA-4.0
film_09	25.00 min	stereo	non-additive	✓	student production	CC-BY-SA-4.0

Audio format & normalization

• FLAC (lossless), 24-bit, 48 kHz throughout — bit-exact to the source PCM (verified maxdiff=0), decoded transparently by soundfile and 🤗 datasets.
• Surround sources are downmixed to stereo via ITU-R BS.775 (Lt = L + 0.707·C + 0.707·Ls, Rt = R + 0.707·C + 0.707·Rs; LFE dropped).
• A single per-production gain is applied uniformly to all files of that production so the loudest signal peaks at −1 dBFS (no clipping in fixed-point). This preserves the additive property (linear_mix = g·(dialogue+music+sfx)) and the relative balance between stems and the production mix. The applied gain is recorded in films.json.
• Additivity verified: max|dialogue+music+sfx − linear_mix| ≈ 2·10⁻⁷ (24-bit quantization floor).

Loading

import soundfile as sf

# additive evaluation: input decomposes exactly into the targets
mix,  sr = sf.read("data/film_02/linear_mix.flac")
dia,  _  = sf.read("data/film_02/dialogue.flac")
mus,  _  = sf.read("data/film_02/music.flac")
sfx,  _  = sf.read("data/film_02/sfx.flac")
# mix == dia + mus + sfx  (to 24-bit precision)

# real-world evaluation: mastered, non-additive input
prod, _  = sf.read("data/film_02/production_mix.flac")

Or via 🤗 datasets (each row = one audio file; filter on component / kind / film_id):

from datasets import load_dataset
ds = load_dataset("audiofolder", data_dir="CineAudioDB", split="test")
stems = ds.filter(lambda r: r["kind"] == "target_stem")

metadata.csv columns: file_name, film_id, component, kind, channels, sample_rate, bit_depth, frames, duration_sec, peak, mix_type, source_type, license.

Processing & provenance

Each production arrived in a different state; all were normalized to the layout above.

• film_01 — 8 professional print-master stem categories (DX, VO, MX, FX, FOL, BG, PFX, ARCH), each as six 24-bit mono channels of a 5.1 mix. Each category was downmixed to stereo (BS.775) and grouped: dialogue = DX+VO, music = MX, sfx = FX+FOL+BG+PFX+ARCH.
• film_02 — stems trimmed of 11.45 s trailing silence to align with the mastered mix; mono center dialogue duplicated to L=R; original 24-bit stereo master used as production_mix.
• film_03 — native mono, 16-bit; kept mono (stereo cannot be honestly recovered) and upcast to 24-bit. The supplied mixture is the mastered production_mix.
• film_04 — reconstructed from a DAW (Ardour) session by rendering each region (source slice, region gain, fades, center pan) into per-category stereo buses; reverb sends and master-bus plugins are intentionally not replicated, which is what makes the stems dry and the mix non-additive. The real session render is the production_mix, advanced by 682 samples (~14 ms) to compensate for master-bus latency and align with the stems.

Known limitations

• film_03 is mono. Its dialogue/music/sfx and both mixes are single-channel.
• film_04 stems are dry reconstructions. They omit master-bus reverb/EQ/limiting; the intro/outro theme music audible in its production_mix is not isolated in the stems (it existed only inside a full-mix bounce and was excluded to keep the music target clean). This is a genuine non-additive gap.
• production_mix ≠ Σ stems for film_02–film_04 by design. Use linear_mix when you need exact additivity.
• Surround productions lose the LFE channel in the stereo downmix (standard practice).

Citation

@inproceedings{cineaudiodb2026,
  title     = {Realistic Data Generation and Evaluation for Cinematic Audio Source Separation},
  author    = {Anonymous},
  booktitle = {Interspeech (under review)},
  year      = {2026}
}

Licensing & attribution

This dataset mixes licenses. See NOTICE.md for the required attributions of the Creative Commons source works. The open-source animations are derived from CC-BY films (© Blender Foundation); the student productions are released under CC-BY-SA 4.0. You must comply with the per-production licenses listed in metadata.csv / films.json.