TTS-AGI/ears-corrupted-dramabox-roundtrip

EARS Corrupted DramaBox Roundtrip

Paired clean/corrupted speech dataset for training speech restoration models. Each sample contains a clean recording from the EARS dataset alongside three corrupted variants produced by the DramaBox neural audio VAE.

Overview

Property	Value
Samples	17,227
Files per sample	5 (4 audio + 1 metadata)
Total audio duration	~96 hours (clean only; ~383 hours across all variants)
Total size	~43 GB
Audio format	MP3, mono, 48 kHz, 256 kbps
Source	CLAPv2/ears_dataset

Data Structure

data/
  p001_162/
    clean.mp3                  # Original EARS recording (ground truth)
    roundtrip.mp3              # DramaBox VAE encode -> decode
    noisy_roundtrip.mp3        # Roundtrip with Gaussian noise injected into latents
    augmented_roundtrip.mp3    # Audio augmented before roundtrip (comb/distortion)
    meta.json                  # Sample metadata
  p001_163/
    ...
  p107_1771/
    ...

Corruption Variants

1. roundtrip.mp3 — Standard VAE Roundtrip

The clean audio is encoded through the DramaBox VAE encoder at 16 kHz, then decoded back to 48 kHz. This produces characteristic neural codec artifacts: spectral smoothing, metallic timbral shifts, and transient smearing.

2. noisy_roundtrip.mp3 — Noisy Latent Roundtrip

Same as above, but Gaussian noise is added to each channel of the VAE latent representation before decoding. The noise magnitude is sampled uniformly between 0.1% and 5% of each channel's standard deviation. This simulates transmission errors or quantization noise in the latent space.

3. augmented_roundtrip.mp3 — Augmented Input Roundtrip

The clean audio is first corrupted with a DSP effect, then passed through the DramaBox VAE roundtrip. This produces compounded artifacts (DSP + codec). The augmentation applied to each sample is recorded in meta.json. Distribution:

Augmentation	Probability	Description
Comb filter	30%	Delay 0.5–4 ms, wet 0.2–0.6
Distortion	30%	Pedalboard Distortion, 3–12 dB drive
Comb + distortion	30%	Sequential comb then distortion
None	10%	Plain roundtrip (same as roundtrip.mp3)

Observed distribution across the dataset: comb 5,235 / distortion 5,053 / comb+distortion 5,179 / none 1,673.

Metadata Format

Each meta.json contains:

{
  "sample_id": "p001_162",
  "text": "When the sunlight strikes raindrops in the air, they act as...",
  "original_sr": 48000,
  "duration_s": 9.18,
  "noise_fraction": 0.0323,
  "augmentation_type": "comb"
}

Field	Description
sample_id	Matches the directory name and original EARS index
text	Transcript from the EARS dataset (truncated to 200 chars)
original_sr	Sample rate of the source EARS recording
duration_s	Duration of the clean audio in seconds
noise_fraction	Fraction of per-channel latent std used as noise magnitude (for noisy_roundtrip.mp3)
augmentation_type	DSP augmentation applied before roundtrip: comb, distortion, comb+distortion, or none

How It Was Made

The dataset was generated using the script process_ears.py with the following pipeline:

1. Streaming: The CLAPv2/ears_dataset is streamed from HuggingFace (never loaded fully into memory).

2. Producer-consumer: A producer thread feeds samples to a multiprocessing queue (maxsize 100). 8 GPU workers consume from the queue in parallel.

3. Per-GPU processing: Each GPU worker loads the DramaBox AudioConditioner (encoder) and AudioDecoder. For each sample:

   • Audio is resampled to 16 kHz, duplicated to stereo, and encoded to VAE latents.
   • Latents are decoded to produce roundtrip.mp3 (48 kHz).
   • Gaussian noise is injected into the latents and decoded for noisy_roundtrip.mp3.
   • The original audio is augmented with comb filter and/or distortion, re-encoded, and decoded for augmented_roundtrip.mp3.
   • The original audio is resampled to 48 kHz and saved as clean.mp3.

4. Output: All audio is saved as mono MP3 at 48 kHz / 256 kbps via ffmpeg.

DramaBox checkpoint: ResembleAI/Dramabox (dramabox-audio-components.safetensors)

Duration Statistics

Statistic	Value
Minimum	0.5 s
Maximum	208.7 s
Mean	20.1 s
Total (clean)	95.5 hours

Usage

from datasets import load_dataset
import torchaudio

# Load a sample
ds = load_dataset("TTS-AGI/ears-corrupted-dramabox-roundtrip")

# Or load directly from disk
clean, sr = torchaudio.load("data/p001_162/clean.mp3")
corrupted, _ = torchaudio.load("data/p001_162/roundtrip.mp3")

This dataset is used to train the Sidon speech restoration vocoder — specifically to remove DramaBox neural codec artifacts while preserving speech quality.

License

CC-BY-4.0 (following the EARS dataset license).