File size: 3,009 Bytes
854ed3f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
# AffectDF

285,797 labelled recordings imported from `/ds-slt/audio/AffectDF`.
The provided protocols define the partitions; source dev maps to validation.

| Table | Rows | Bonafide | Spoof |
|---|---:|---:|---:|
| `data.parquet` | 285,797 | 19,108 | 266,689 |
| `parquets/train.parquet` | 86,999 | 6,000 | 80,999 |
| `parquets/val.parquet` | 23,330 | 3,000 | 20,330 |
| `parquets/test.parquet` | 175,468 | 10,108 | 165,360 |

`data.parquet` retains all supplied protocol rows and metadata. Its three
partitions are in `parquets/`. Every audio `path` points inside this dataset's
`wav/` tree; `original_path` records the original location for provenance.
Audio files are independent local copies in SDD.

## Metadata

`ID` is unique within the dataset and includes the language and source partition.
`original_ID` preserves the source waveform stem. `label` uses bonafide/spoof;
`source_label` preserves the original label. `dataset_name`, `subset`, `language`,
and `split` retain the dataset, original subset, language, and canonical partition.

AffectDF also preserves `speaker_id`, `attack_id`, `emotion`, `generation_type`,
and `generation_model` exactly from its seven-field protocols. Source label
`real` maps to `bonafide`. Audio lives in `wav/Train/`, `wav/Dev/`, and `wav/Test/`.
Original protocols are in `protocols/`. The source documentation and stated
CC BY-NC 4.0 license are retained in [SOURCE_README.md](SOURCE_README.md).

Some original local audio was owner-only and unreadable. The audio was
recovered from the [public release](https://huggingface.co/datasets/AffectDF/AffectDF_EmotionSDD) at revision
`529ed470208e52e59c911a531cf803f8e42c8755`. All three public protocols match the local protocol
rows. Every recovered filename and size was checked, and 30 readable source
recordings were compared byte-for-byte. `original_path` remains a provenance
field; training uses the accessible SDD `path`. Full evidence is in
`/ds-slt/audio/sdd-research-baseline/reports/affectdf-public-recovery.json`.

## Verification and training

Every protocol row was checked against the copied audio inventory. Class-based
full-waveform samples were decoded for each source subset. A subsequent audit
checked every parquet row, membership of each partition in the full table,
canonical split names, audio path prefixes, split path overlap, and sampled
audio reads. These checks are recorded in
`/ds-slt/audio/sdd-research-baseline/reports/emotional-datasets-audit.json`.
They do not constitute an exhaustive decode of every recording.

`import.json` records source protocol SHA-256 hashes, exact counts, copy details,
source-ID overlap checks, and any audio without protocol metadata.

Separate AASIST and MLP experiments use this dataset's own train/validation
partitions. Test data stays held out. Configs and commands are in
`/ds-slt/audio/yelkheir/DeepFense/deepfense/config/experiments/EmotionalSDD/README.md`.
The handover [runbook](../../sdd-research-baseline/README.md) records their Slurm jobs.