Datasets:

Nanboy
/

RVCBench

Cloned Repository:
- Source: https://github.com/wspr-ncsu/robocall-audio-dataset
- Contains 1,432 real robocall audio recordings from FTC's Project Point of No Entry
- Includes transcripts for all recordings
Organized by Scam Type:
- Created Different_spamtype/ folder with 11 categories
- Each category contains 5 audio files with transcripts
- Filtered for complete sentences only (removed fragments)
- Created corresponding .txt files for each audio
Scam Categories:
- Amazon_Scam (5 files)
- SSA_Scam (Social Security) (5 files)
- Auto_Warranty_Scam (5 files)
- IRS_Tax_Scam (5 files)
- Utility_Scam (5 files)
- Telecom_Scam (5 files)
- Credit_Card_Scam (5 files)
- CBP_Scam (Customs & Border Protection) (5 files)
- Student_Loan_Scam (5 files)
- Package_Delivery_Scam (5 files)
- CashApp_Scam (2 files)

Step 2: VCTK-Based Robocall Dataset

Location: /content/drive/MyDrive/oct25/AudioWatermarkBench/data/audiobenchmarkdataset/robotcall/

Selected 10 VCTK Speakers:
- Random selection: p227, p232, p234, p237, p238, p251, p253, p262, p283, p288
- All speakers use the same scam texts for consistency
Created Audio-Based Entries:
- 2 ori_pth audios per speaker (VCTK clean speech)
- 5 gt_pth audios per speaker (VCTK clean speech)
- Randomly paired gt to ori
- Total: 50 audio-based entries (5 per speaker)
Added Scam Text Entries:
- 2 texts from each of the 11 scam categories
- Same ori_pth (VCTK audio) but gt_text is robocall scam
- gt_pth contains scam category name (not blank)
- Total: 220 text-only entries (22 per speaker)

Dataset Structure

Directory Layout

robotcall/
├── audios/
│   ├── p227/  (7 .wav files: 2 ori + 5 gt)
│   ├── p232/  (7 .wav files)
│   ├── p234/  (7 .wav files)
│   ├── p237/  (7 .wav files)
│   ├── p238/  (7 .wav files)
│   ├── p251/  (7 .wav files)
│   ├── p253/  (7 .wav files)
│   ├── p262/  (7 .wav files)
│   ├── p283/  (7 .wav files)
│   └── p288/  (7 .wav files)
└── filelists/
    ├── p227.json  (27 entries: 5 audio + 22 text)
    ├── p232.json  (27 entries)
    ├── p234.json  (27 entries)
    ├── p237.json  (27 entries)
    ├── p238.json  (27 entries)
    ├── p251.json  (27 entries)
    ├── p253.json  (27 entries)
    ├── p262.json  (27 entries)
    ├── p283.json  (27 entries)
    ├── p288.json  (27 entries)
    └── all_speakers.json  (270 entries combined)

Statistics

Total Speakers: 10
Total Entries: 270
- Audio-based: 50 (5 per speaker)
- Text-only (scam): 220 (22 per speaker)
Audio Files: 70 total (7 per speaker)
Scam Texts per Category: 2 (consistent across all speakers)

JSON Entry Format

Entry Type 1: Audio-Based (VCTK to VCTK)

{
  "ori_pth": "audios/p227/p227_397_mic1.wav",
  "ori_spk": "p227",
  "ori_lang": "EN",
  "ori_text": "They have not got anyone.",
  "gt_pth": "audios/p227/p227_322_mic1.wav",
  "gt_spk": "p227",
  "gt_text": "The judge was really nice.",
  "ori_phonemes": "",
  "ori_tone": "",
  "ori_word2ph": "",
  "gt_phonemes": "",
  "gt_tone": "",
  "gt_word2ph": ""
}

Entry Type 2: Text-Only (VCTK to Robocall Scam)

{
  "ori_pth": "audios/p227/p227_397_mic1.wav",
  "ori_spk": "p227",
  "ori_lang": "EN",
  "ori_text": "They have not got anyone.",
  "gt_pth": "Amazon_Scam",
  "gt_spk": "p227",
  "gt_text": "We would like to inform you that there is an order placed for Apple iPhone 11 Pro using your Amazon account. If you do not authorize this order, press 1 or press 2 to authorize this order.",
  "ori_phonemes": "",
  "ori_tone": "",
  "ori_word2ph": "",
  "gt_phonemes": "",
  "gt_tone": "",
  "gt_word2ph": ""
}

Field Descriptions

Field	Description
`ori_pth`	Original audio file path (always VCTK clean speech)
`ori_spk`	Original speaker ID (e.g., 'p227')
`ori_lang`	Original language (always 'EN')
`ori_text`	Original transcript (VCTK clean speech text)
`gt_pth`	Ground truth path - Audio entries: VCTK audio file path - Text entries: Scam category name (e.g., 'Amazon_Scam')
`gt_spk`	Ground truth speaker ID
`gt_text`	Ground truth text - Audio entries: VCTK transcript - Text entries: Robocall scam transcript
`ori_phonemes`	Phoneme sequence (kept blank)
`ori_tone`	Tone sequence (kept blank)
`ori_word2ph`	Word-to-phoneme mapping (kept blank)
`gt_phonemes`	Phoneme sequence (kept blank)
`gt_tone`	Tone sequence (kept blank)
`gt_word2ph`	Word-to-phoneme mapping (kept blank)

Use Cases

Voice Conversion:
- Convert clean VCTK speech to robocall scam content
- Maintain speaker identity while changing text
Audio Watermarking:
- Embed watermarks in synthesized robocall audio
- Test watermark robustness against voice conversion
Scam Detection:
- Train models to detect scam content patterns
- Compare clean vs. scam speech characteristics
Text-to-Speech Evaluation:
- Generate robocall audio from scam texts
- Evaluate naturalness and intelligibility

Key Design Decisions

Complete Sentences Only:
- All robocall transcripts start with complete sentences
- No mid-sentence fragments
Consistent Scam Texts:
- All speakers use the same 2 texts per scam category
- Enables controlled comparison across speakers
Scam Type in gt_pth:
- For text-only entries, gt_pth contains category name
- Makes it easy to identify scam type without parsing text
- Maintains consistent JSON structure
Random Pairing:
- GT texts randomly assigned to ori audios
- Provides variety in ori-gt combinations

Scripts Used

organize_scams.py - Organized robocall dataset by scam type with complete sentences
create_robocall_dataset.py - Created initial VCTK-based dataset with audio entries
add_robocall_texts.py - Added scam text entries with scam category in gt_pth

Notes

All audio files are from VCTK corpus (16kHz, mono)
Robocall transcripts are from real FTC-investigated scam calls
Dataset designed for research purposes only
All scam texts are complete, coherent sentences suitable for TTS