CineAudioSynth

Synthetic cinematic audio for source separation. 453 scenes, ~22.8 h, 48 kHz / 16-bit / stereo WAV.

Each data/scene_NNNN/ contains:

• linear/ — additive render: mix.wav is the BIT-EXACT 16-bit sum of the four stems (speech, music, ambience, sfx) — max |mix − Σstems| = 0, verified per scene. Includes gain_envelope.json (sidechain ducking envelopes).
• release/ — mastered render of the same scene (compression/limiting/loudness on the mix bus; intentionally non-additive).
• metadata.json — scene type, DSP/critic settings, SFX event timestamps and descriptions, loudness metrics, additivity measurements.

Directory structure

CineAudioSynth/
├── README.md
├── ATTRIBUTIONS.md          # per-corpus & per-clip credits + citations
├── LICENSE
└── data/
    ├── scene_0001/
    │   ├── linear/                 # additive render
    │   │   ├── speech.wav
    │   │   ├── music.wav
    │   │   ├── ambience.wav
    │   │   ├── sfx.wav
    │   │   ├── mix.wav             # == speech + music + ambience + sfx  (bit-exact)
    │   │   └── gain_envelope.json  # sidechain ducking envelopes
    │   ├── release/                # mastered render (non-additive)
    │   │   ├── speech.wav  music.wav  ambience.wav  sfx.wav
    │   │   └── mix.wav
    │   └── metadata.json           # scene plan, DSP settings, loudness, additivity
    ├── scene_0002/
    └── ...                         # 453 scenes (scene_0001 … scene_0453)

Within a scene, every WAV shares the same length, sample rate (48 kHz), bit depth (16-bit) and channel count (stereo).

Quick start

# the whole dataset (~158 GB) ...
from huggingface_hub import snapshot_download
root = snapshot_download("disco-eth/cineaudiosynth", repo_type="dataset")

# ... or grab a single scene
from huggingface_hub import hf_hub_download
for f in ["linear/speech.wav", "linear/music.wav", "linear/ambience.wav",
          "linear/sfx.wav", "linear/mix.wav", "release/mix.wav", "metadata.json"]:
    hf_hub_download("disco-eth/cineaudiosynth", f"data/scene_0001/{f}", repo_type="dataset")

Load with 🤗 datasets

A root metadata.csv ships an audiofolder manifest — one row per WAV with scene-level columns — so the dataset loads directly:

from datasets import load_dataset

ds = load_dataset("disco-eth/cineaudiosynth", split="train")                  # decodes audio (needs `torchcodec`)
ds = load_dataset("disco-eth/cineaudiosynth", split="train", streaming=True)  # stream, no 158 GB download

# filter to what you need — e.g. linear-render stems of dialogue-heavy scenes
stems = ds.filter(lambda r: r["render"] == "linear" and not r["is_mix"]
                            and r["scene_type"] == "dialogue_heavy")

ex = next(iter(stems))
ex["audio"]["array"], ex["audio"]["sampling_rate"], ex["scene_id"], ex["component"]

Decoding uses torchcodec (pip install torchcodec); to read paths/metadata only, add .cast_column("audio", datasets.Audio(decode=False)).

metadata.csv columns: file_name, scene_id, render, component, is_mix, scene_type, has_music, has_ambience, n_sfx_events, music_tag, ambience_cue, linear_additivity_verified.

Loading a scene

import soundfile as sf
scene = "data/scene_0001"

# linear (additive) render — mix is the bit-exact 16-bit sum of the four stems
speech,     sr = sf.read(f"{scene}/linear/speech.wav")
music,      _  = sf.read(f"{scene}/linear/music.wav")
ambience,   _  = sf.read(f"{scene}/linear/ambience.wav")
sfx,        _  = sf.read(f"{scene}/linear/sfx.wav")
linear_mix, _  = sf.read(f"{scene}/linear/mix.wav")
# linear_mix == speech + music + ambience + sfx   (max abs diff == 0)

# release render — same scene, mastered on the mix bus (intentionally non-additive)
release_mix, _ = sf.read(f"{scene}/release/mix.wav")

Verify additivity

import numpy as np, soundfile as sf
lin = "data/scene_0001/linear"
stems = sum(sf.read(f"{lin}/{s}.wav")[0] for s in ["speech", "music", "ambience", "sfx"])
mix, _ = sf.read(f"{lin}/mix.wav")
print("max |mix - Σstems|:", np.max(np.abs(mix - stems)))   # -> 0.0

Iterate every scene

import glob, os, soundfile as sf

root = "data"  # or the path returned by snapshot_download(...)
for scene in sorted(glob.glob(os.path.join(root, "scene_*"))):
    mix, sr = sf.read(os.path.join(scene, "linear", "mix.wav"))
    targets = {s: sf.read(os.path.join(scene, "linear", f"{s}.wav"))[0]
               for s in ["speech", "music", "ambience", "sfx"]}
    # ... train / evaluate your separator on (mix, targets)

Reading scene metadata

import json
meta = json.load(open("data/scene_0001/metadata.json"))

meta["scene_type"]                     # e.g. "balanced"
meta["linear_additivity_verified"]     # True
meta["scene_plan"]["events"]           # [{"timestamp": 3.0, "description": "..."}, ...]
meta["outputs"]["linear"]["loudness"]  # integrated LUFS / true-peak dBTP

linear/gain_envelope.json holds the per-stem, speech-triggered sidechain-ducking keyframes used to build the linear mix.

License & attribution

This dataset is released under CC BY-NC-SA 4.0 (non-commercial, share-alike). Individual stems remain governed by the licenses of their source material; music and sound-effect assets appear only within mixed scenes and may not be extracted and re-published as standalone collections. Full per-track and per-clip credits, plus citations for all source corpora, are in ATTRIBUTIONS.md.

Source corpora and asset libraries used by the CineAudioGen pipeline:

Source	Content	License
Expresso	speech	CC BY-NC 4.0
RAVDESS	speech	CC BY-NC-SA 4.0
ASED	speech	per its distribution page (cite authors)
NonverbalTTS	speech	Apache-2.0 (annotations); audio CC BY / CC BY-NC
FSD50K	SFX & ambience	CC BY 4.0; per-clip CC0 / CC BY / CC BY-NC / CC Sampling+
FMA	music	per-track CC licenses
Chosic	music	per-track (CC and artist terms; attribution required)
no-copyright-music.com	music	royalty-free (provider terms; attribution required)

When using this dataset, cite the source corpora as listed in ATTRIBUTIONS.md.

Research use only. Not for commercial use. Derivatives must be shared under the same license (CC BY-NC-SA 4.0).