yoona-J/BAC

Band-wise Acoustic Conditioning Dataset

This repository contains the preprocessed data structure and derived acoustic features used for the experiments in Multi-head Acoustic Conditioning for Low-information Conversational ASR.

The original speech corpus was provided by the National Institute of Korean Language (NIKL)Link. This repository is intended to document and distribute the processed artifacts used for reproducibility.

Code Repository

---

English

Dataset Overview

This dataset is organized for reproducing the ASR baseline, Band-wise Acoustic Conditioning (BAC), and feature-group ablation experiments.

The released artifacts are divided into four groups:

raw/
hallu_acoustic_feats_fix/
hallu_acoustic_feats_temporal_fix/
hallu_acoustic_feats_voice/

1. Raw Speech Corpus

• Raw Dataset download

The preprocessed speech corpus is organized under raw/org with separate train, dev, and test splits.

raw/
└── org/
    ├── train/
    ├── dev/
    └── test/
        ├── audio_format
        ├── data/
        ├── feats_type
        ├── logs/
        ├── spk2utt
        ├── text
        ├── utt2num_samples
        ├── utt2spk
        └── wav.scp

Each split follows the ESPnet-style data directory format.

• wav.scp: maps utterance IDs to audio sources or loading commands.
• text: contains the reference transcription for each utterance.
• utt2spk: maps utterance IDs to speaker IDs.
• spk2utt: maps speaker IDs to their corresponding utterances.
• utt2num_samples: stores the number of audio samples for each utterance.
• audio_format: specifies the audio format.
• feats_type: specifies the ESPnet feature type.
• logs/: contains preprocessing-related logs.

2. Full Acoustic Features for BAC

• Acoustic Features Dataset download

hallu_acoustic_feats_fix contains the full utterance-level acoustic descriptor used for BAC training and evaluation.

hallu_acoustic_feats_fix/
├── acoustic_stats.json
├── train/
├── dev/
├── test/
│   ├── acoustic_feats.scp
│   ├── acoustic_feats_shuffled_seed42.scp
│   ├── norm/
│   └── raw/
└── test_1k/

The BAC acoustic descriptor consists of nine features.

Temporal features

• silence ratio
• number of pauses
• mean pause duration

Voice / prosodic features

• F0 standard deviation
• F0 range
• RMS energy
• HNR
• jitter
• shimmer

Main files:

• acoustic_stats.json: normalization statistics computed from the training split
• acoustic_feats.scp: mapping between utterance IDs and acoustic feature files
• raw/: acoustic features before normalization
• norm/: acoustic features normalized using training-set statistics
• test_1k/: 1,000-utterance test subset

3. Temporal Acoustic Features

• Temporal Acoustic Features Dataset download

hallu_acoustic_feats_temporal_fix contains only the temporal features used for the temporal ablation experiment.

hallu_acoustic_feats_temporal_fix/
├── train/
├── dev/
├── test/
│   ├── acoustic_feats.scp
│   └── norm/
└── test_1k/

Included features:

• silence ratio
• number of pauses
• mean pause duration

4. Voice / Prosodic Acoustic Features

• Voice Acoustic Features Dataset download

hallu_acoustic_feats_voice contains only the voice/prosodic features used for the voice/prosodic ablation experiment.

hallu_acoustic_feats_voice/
├── train/
├── dev/
├── test/
│   ├── acoustic_feats.scp
│   └── norm/
└── test_1k/

Included features:

• F0 standard deviation
• F0 range
• RMS energy
• HNR
• jitter
• shimmer

---

Notes

• The dataset structure follows the files used in the experiments.
• train, dev, and test correspond to the training, validation, and test splits.
• test_1k contains the 1,000-utterance subset used in selected analyses.
• The original corpus was provided by the National Institute of Korean Language (NIKL).

---

Korean

Dataset Overview

본 데이터는 ASR baseline, Band-wise Acoustic Conditioning (BAC), 그리고 feature-group ablation 실험의 재현을 위해 구성되었습니다.

데이터는 다음 네 가지 구성으로 나뉩니다.

raw/
hallu_acoustic_feats_fix/
hallu_acoustic_feats_temporal_fix/
hallu_acoustic_feats_voice/

1. Raw Speech Corpus

• Raw Dataset download

전처리된 음성 데이터는 raw/org 아래에 train, dev, test split으로 구성되어 있습니다.

raw/
└── org/
    ├── train/
    ├── dev/
    └── test/
        ├── audio_format
        ├── data/
        ├── feats_type
        ├── logs/
        ├── spk2utt
        ├── text
        ├── utt2num_samples
        ├── utt2spk
        └── wav.scp

각 split은 ESPnet-style data directory 형식을 따릅니다.

• wav.scp: utterance ID와 audio source 또는 loading command를 매핑합니다.
• text: 각 utterance의 reference transcription을 포함합니다.
• utt2spk: utterance ID와 speaker ID를 매핑합니다.
• spk2utt: speaker ID와 해당 utterance 목록을 매핑합니다.
• utt2num_samples: 각 utterance의 audio sample 수를 저장합니다.
• audio_format: audio format 정보를 저장합니다.
• feats_type: ESPnet feature type 정보를 저장합니다.
• logs/: 전처리 관련 로그를 포함합니다.

2. Full Acoustic Features for BAC

• Acoustic Features Dataset download

hallu_acoustic_feats_fix는 BAC 학습 및 평가에 사용되는 전체 utterance-level acoustic descriptor를 포함합니다.

hallu_acoustic_feats_fix/
├── acoustic_stats.json
├── train/
├── dev/
├── test/
│   ├── acoustic_feats.scp
│   ├── acoustic_feats_shuffled_seed42.scp
│   ├── norm/
│   └── raw/
└── test_1k/

BAC acoustic descriptor는 총 9개 feature로 구성됩니다.

Temporal features

• silence ratio
• number of pauses
• mean pause duration

Voice / prosodic features

• F0 standard deviation
• F0 range
• RMS energy
• HNR
• jitter
• shimmer

주요 파일:

• acoustic_stats.json: training split을 기준으로 계산된 normalization statistics
• acoustic_feats.scp: utterance ID와 acoustic feature 파일의 mapping
• raw/: normalization 이전의 acoustic features
• norm/: training statistics로 normalization된 acoustic features
• test_1k/: 1,000-utterance test subset

3. Temporal Acoustic Features

• Temporal Acoustic Features Dataset download

hallu_acoustic_feats_temporal_fix는 temporal feature만을 사용하는 ablation 실험용 데이터입니다.

hallu_acoustic_feats_temporal_fix/
├── train/
├── dev/
├── test/
│   ├── acoustic_feats.scp
│   └── norm/
└── test_1k/

포함 feature:

• silence ratio
• number of pauses
• mean pause duration

4. Voice / Prosodic Acoustic Features

• Voice Acoustic Features Dataset download

hallu_acoustic_feats_voice는 voice/prosodic feature만을 사용하는 ablation 실험용 데이터입니다.

hallu_acoustic_feats_voice/
├── train/
├── dev/
├── test/
│   ├── acoustic_feats.scp
│   └── norm/
└── test_1k/

포함 feature:

• F0 standard deviation
• F0 range
• RMS energy
• HNR
• jitter
• shimmer