Skip to content

Latest commit

 

History

History
221 lines (166 loc) · 5.32 KB

DatasetCard.md

File metadata and controls

221 lines (166 loc) · 5.32 KB

Callytics Speaker Verification Dataset (CSVD)

This dataset, named Callytics Speaker Verification Dataset (CSVD), was created to test and develop Speaker Verification models specifically for the Callytics project. It primarily features Turkish speakers and was developed in collaboration with Turkish student volunteers. The recordings originate from WhatsApp phone calls, which were captured using the Cube ASR mobile application in AMR format and later converted to WAV using FFmpeg.

Table of Contents

Features

  • Dataset Name: Callytics Speaker Verification Dataset (CSVD)
  • Version: v1 (Initial)
  • Purpose: This dataset is created to test and develop Speaker Verification models, especially for the Callytics project.
  • Language / Accent: Primarily Turkish speakers
  • Contributors: Created in collaboration with Turkish student volunteers.
  • Recording Source: WhatsApp calls (originally AMR) captured Cube ASR mobile application and then converted to WAV using FFmpeg.

This version (v1) contains 10 samples of phone calls involving Customer Service Representative (CSR) and * Customer roles. The audio files are stored in WAV format with the following specifications:

  • Channels: Mono
  • Format: WAV
  • Audio Bitrate: 128 kbps
  • Sample Rate: 8.00 kHz
  • Conversion: AMR → WAV

File Structure

.
└── speakerverification
    ├── call
    │   └── v1
    │       ├── Call01.wav
    │       ├── Call02.wav
    │       └── metadata.json
    ├── DatasetCard.md
    ├── LICENSE
    ├── README.md
    └── voiceprint
        └── v1
            ├── CSR01.wav
            └── metadata.json

6 directories, 7 files
  • call: Contains call samples (Call01.wav, Call02.wav, etc.) plus a metadata.json.
  • voiceprint/v1: Contains shorter CSR recordings (CSR01.wav, etc.) plus a metadata.json.
  • DatasetCard.md: This document, explaining dataset details.
  • LICENSE: License file for the dataset.

Metadata

Call Metadata
Sample
  {
    "file_id": "Call01.wav",
    "csr_id": "CSR01",
    "duration": 1.13,
    "recorded": "2025-02-11"
  }
  • file_id: Filename of the audio file.
  • csr_id: CSR identifier.
  • duration: Duration of the call.
  • recorded: Recording date.
Voiceprint Metadata
Sample
  {
    "file_id": "CSR01.wav",
    "csr_id": "CSR01",
    "gender": "F",
    "age": 22,
    "duration": 0.7,
    "transcript": "Hello and welcome to Callytics! My name is Emily. How may I assist you?",
    "recorded": "2025-02-11"
  }
  • file_id: Filename of the audio file.
  • csr_id: CSR identifier.
  • gender: Gender of CSR.
  • age: Age of CSR.
  • duration: Duration of the Voice Print.
  • recorded: Recording date.

Transcript

We provide SRT transcripts for each call under the call/v1 directory. Each .srt file contains:

  • Speaker labels
  • Time stamps in HH:MM:SS,MMM format
  • Conversation text

Such as:

  • Call01.wav has a corresponding Call01.srt
  • Call02.wav has a corresponding Call02.srt
Annotation Process

The timestamps in the transcripts were manually annotated using Audacity for precise time labeling.

Transcript Sample
1
00:00:03,453 --> 00:00:09,166
[CSR] Hello and welcome to Callytics! My name is Emily. How may I assist you?

2
00:00:09,582 --> 00:00:20,410
[Customer] Hi, Emily. I’ve been having trouble receiving notification emails.
They used to arrive instantly, but now there’s a long delay.

3
00:00:20,599 --> 00:00:27,879
[CSR] I’m sorry for the inconvenience. Let me take a look.
Could you give me your account details so I can verify the email settings?

Versioning

  • v1 (Initial)

Upcoming

  • Potentially larger sets of calls
  • Inclusion of new languages or dialects
  • Measure Audio Data Quality
    • General Data Quality Indicators
      • Completeness
      • Accuracy
      • Consistency
      • Uniqueness (Duplicate Check)
      • Coverage / Diversity
      • Timeliness
      • Class Balance and Metadata Richness
    • Audio-Specific Quality Indicators
      • Signal-to-Noise Ratio (SNR)
      • Distortion Measures
      • Clipping Rate (Clipped Samples)
      • Silence Ratio
      • Noise Types and Levels
      • Reverberation (Echo) Ratio / DRR (Direct to Reverberant Ratio)
      • Speaker / Content Distribution
      • Transcription Quality

Licence

Team

Contact

Citation

@software{       CSVD,
  author       = {Bunyamin Ergen},
  title        = {Callytics Speaker Verification Dataset},
  year         = {2025},
  month        = {02},
  url          = {https://github.com/bunyaminergen/Callytics},
  version      = {v1},
}