File size: 3,099 Bytes
840242d
 
 
 
 
 
 
 
f4d634d
 
 
840242d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
af89b30
840242d
 
 
 
af89b30
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
---
license: cc-by-nc-4.0
language:
- en
size_categories:
- 10K<n<100K
---

[ArXiv Paper Publication Here: "Real-World En Call Center Transcripts Dataset with PII Redaction"](https://arxiv.org/abs/2507.02958)


This dataset includes **91,706 high-quality transcriptions** corresponding to approximately **10,500 hours** of **real-world call center conversations** in **English**, collected across various industries and global regions. The dataset features both **inbound and outbound** calls and spans multiple accents, including **Indian**, **American**, and **Filipino** English. All transcripts have been **carefully redacted for PII** and enriched with **word-level timestamps** and **ASR confidence scores**, making it ideal for training robust speech and language models in real-world scenarios.

* 🗣️ **Language & Accents**: English (Indian, American, Filipino)
* 📞 **Call Types**: Inbound and outbound customer service conversations
* 🏢 **Source**: Sourced via partnerships with BPO centers across a range of industries
* 🔊 **Audio Length**: 10,500+ hours of corresponding real-world audio (not included in this release)
* 📄 **Transcripts**: 91,706 JSON-formatted files with:

  * Word-level timestamps
  * ASR confidence scores
  * Categorized by domain, topic, and accent
  * Redacted for privacy

🔧 **Processing Pipeline**:

1. Raw, uncompressed audio was downloaded directly from BPO partners to maintain acoustic integrity.

2. Calls were tagged by **domain**, **accent**, and **topic** (inbound vs outbound).

3. Transcription was done using **AssemblyAI’s paid ASR model**.

4. Transcripts and audios were **redacted for PII** based on the following list:

   ```
   account_number, banking_information, blood_type, credit_card_number, credit_card_expiration, 
   credit_card_cvv, date, date_interval, date_of_birth, drivers_license, drug, duration, 
   email_address, event, filename, gender_sexuality, healthcare_number, injury, ip_address, 
   language, location, marital_status, medical_condition, medical_process, money_amount, 
   nationality, number_sequence, occupation, organization, passport_number, password, person_age, 
   person_name, phone_number, physical_attribute, political_affiliation, religion, statistics, 
   time, url, us_social_security_number, username, vehicle_id, zodiac_sign
   ```

5. A manually QA’d subset was used to calculate **word error rate (WER)**, with the overall transcription accuracy estimated at **96.131%**.

6. Final output is provided in **JSON format**, with cleaned and standardized fields.

📜 **Paper Coming Soon**: A detailed paper describing the full pipeline, challenges, and benchmarks is now published here: https://arxiv.org/abs/2507.02958
📣 **Want Updates?** Drop a comment in the **community section** to be notified when the paper goes live.

🔐 **License**: Provided **strictly for research and AI model development**. **Commercial use, resale, or redistribution is prohibited.**

🎓 **Brought to you by AIxBlock and 3 independent researchers: Gaurav Chawla, Raghu Banda, Caleb DeLeeuw **