International Robocalls Dataset

International Robocalls Dataset Dataset accompanying the paper:Robocalls: A Worldwide or US-only Problem? Analyzing Spam and Fraud in International Phone Calls Paper can be found here: https://doi.org/10.48550/arXiv.2606.31790 Overview Calls were collected via a honeypot-based approach; phone numbers were configured to deliberately answer all incoming calls. This dataset is the first publicly available multilingual and international robocall dataset. The dataset spans three modalities: call metadata (call detail records - CDRs), audio recordings, and ASR transcripts, covering both US-domestic and international robocall traffic. Dataset Structure data_processed/├── metadata/│ └── cdrs.csv # 8,747,351 anonymized call records├── audio/│ ├── audio_files.csv # Index of all audio files (language, region, filename)│ ├── english_us/ # 259 files│ ├── english_int/ # 242 files│ ├── spanish_us/ # 39 files│ ├── spanish_int/ # 45 files│ ├── chinese_us/ # 36 files│ └── polish_int/ # 56 files└── text/ ├── us.csv # 724 US robocall transcripts (22 clusters) └── int.csv # 115 international robocall transcripts (6 clusters) Modalities 1. Metadata (metadata/cdrs.csv) Anonymized Call Detail Records (CDRs) for 8,747,351 calls. Column Description caller_id Anonymized numeric caller identifier callee_id Anonymized numeric callee (honeypot) identifier timestamp Unix timestamp in milliseconds (UTC) caller_country ISO 3166-1 alpha-2 country code of the caller callee_country ISO 3166-1 alpha-2 country code of the callee caller_state US state abbreviation (US callers only; empty otherwise) callee_state US state abbreviation (US callees only; empty otherwise) local_datetime ISO 8601 local datetime with UTC offset Top caller countries: US (85.7%), NG (3.9%), GB (2.4%), SA (0.75%), CA (0.71%), PL (0.64%), KE (0.59%) Top callee countries: US (95.0%), GB (2.5%), PL (0.67%), CA (0.38%), DK (0.36%), PH (0.36%) Caller and callee phone numbers have been replaced with consistent anonymous numeric identifiers. The same caller appearing in multiple records retains the same caller_id, preserving campaign-level structure while removing personally identifiable information. 2. Audio (audio/) 677 WAV audio recordings of verified robocalls, all confirmed by at least one human annotator. Format: WAV, 8 kHz, mono, 16-bit PCM (standard telephony quality) Index file (audio_files.csv): Column Description language Language of the call (english, spanish, chinese, polish) region Origin of the honeypot (us = US-domestic, int = international) filename Filename within the corresponding subdirectory Breakdown by language and region: Subfolder Language Region Number of files english_us English US 259 english_int English International 242 polish_int Polish International 56 spanish_int Spanish International 45 spanish_us Spanish US 39 chinese_us Chinese (Mandarin) US 36 Total / / 677 3. Text Transcripts (text/) ASR transcripts generated using Whisper large-v3-turbo, for calls where sufficient speech was detected (via SileroVAD). Transcripts are grouped by robocall campaign - a cluster of calls sharing the same script with minor variations (e.g., different spoofed caller names or phone numbers). Schema (both us.csv and int.csv): Column Description cluster_id Integer ID identifying the robocall campaign cluster campaign Assigned label for the campaign type (e.g., ROBOCALL: IRS/Tax Scam) transcript Full ASR transcript of the call Personal names and phone numbers appearing in transcripts have been replaced with `[NAME]` and `[PHONE]` placeholders respectively. US transcripts (us.csv): 724 records across 22 clusters Campaign categories present: Bank Fraud Alert Scam Benefits/Grants Scam Education/Student Loan Scam Financial Spam (Generic) Google Business Listing Scam Health Insurance Scam Insurance Quote Spam IRS/Tax Scam Loan/Credit Scam Political Campaign Telemarketing (Generic) Vehicle Warranty Scam Unclassified International transcripts (int.csv): 115 records across 6 clusters Campaign categories present: Financial Spam (Generic) Legal Threat Scam Loan/Debt Scam (Hindi) Telemarketing (Generic) Unclassified Data Collection Calls were collected using a honeypot infrastructure: phone numbers with no publicly advertised purpose were provisioned to automatically answer all incoming calls and record the audio. Voice activity detection (SileroVAD) was applied to filter recordings with no spoken content. Calls under 5 seconds of total duration were discarded prior to further processing. The audio subset published here represents a curated sample of calls that were manually verified as robocalls by a human annotator. The metadata (CDRs) covers the full dataset including calls without audio recordings, provided that the caller hade made at least 2 calls towards the honeypot. Privacy and Anonymization Phone numbers (caller and callee) have been replaced with consistent anonymous integer IDs. Mapping from IDs to real numbers is not published. Names and phone numbers within transcripts have been replaced with `[NAME]` and `[PHONE]` placeholders. All callee numbers belong to the honeypot infrastructure controlled by the authors; no third-party user data is exposed. Geographic granularity is limited to country level (and US state level where available). License This dataset is released under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) You are free to share and adapt the material for any non-commercial use, provided appropriate credit is given. Commercial use: contact for a separate license. Citation If you use this dataset, please cite: Altwlkany, K., Merćep, A., Đuričić, T., Kapetanović, A., and Lacic, E. (2026). Robocalls: A Worldwide or US-only Problem? Analyzing Spam and Fraud in International Phone Calls. arXiv preprint arXiv:2606.31790. Bibtex: @misc{altwlkany2026robocalls, title={Robocalls: A Worldwide or US-only Problem? Analyzing Spam and Fraud in International Phone Calls}, author={Kemal Altwlkany and Andro Merćep and Tomislav Đuričić and Ante Kapetanović and Emanuel Lacic}, year={2026}, eprint={2606.31790}, archivePrefix={arXiv}, primaryClass={cs.CR}, url={https://arxiv.org/abs/2606.31790}, } Contact Feel free to contact any of the authors via: [name.surname]@infobip.com