tiagogarrel/92k-real-world-call-center-scripts-english

ArXiv Paper Publication Here: "Real-World En Call Center Transcripts Dataset with PII Redaction"

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 **