atiyehghm/shemo_persian_diarization_dataset

Dataset Description

The shemo_persian_diarization_dataset is a synthetic multi-speaker speech dataset designed for training and evaluating speaker diarization models in the Persian (Farsi) language. It is built using utterances from the Shemo dataset and processed through a sophisticated synthesis framework to simulate realistic conversational dynamics.

• Language: Persian (Farsi)
• Total Duration: ~70 Hours
• Number of Records: 5,000 audio tracks
• Format: 16kHz Mono WAV with standard RTTM annotations

Dataset Summary

Metric	Value
Total Audio Files	5,000
Total Duration	70 Hours
Sample Rate	16,000 Hz
Annotation Format	RTTM (NIST standard)
Primary Language	Persian

Conversation Types

The dataset is categorized into three distinct interaction styles to ensure model robustness:

• Dialog (2,497 tracks): Multi-speaker turn-taking with natural pauses.
• Overlap (2,000 tracks): Competitive speech where multiple speakers talk simultaneously.
• Monologue (503 tracks): Extended speech from a single speaker.

Speaker Distribution

The tracks vary in complexity, featuring between 1 and 6 distinct speakers:

• 1 Speaker: 503 tracks
• 2 Speakers: 876 tracks
• 3 Speakers: 909 tracks
• 4 Speakers: 919 tracks
• 5 Speakers: 943 tracks
• 6 Speakers: 850 tracks

Generation Methodology

This dataset was generated using a controlled synthetic framework designed to mimic real-world acoustic environments.

1. Source Data

Individual utterances were extracted from the Shemo Persian emotional speech corpus. Each speaker's audio was validated for minimum duration and sample count to ensure high-quality synthesis.

2. Composition Algorithm

For every track, the system:

• Selects a Dialog Type: Randomly assigns the track as a monologue, dialog, or overlap scenario.
• Speaker & Utterance Selection: Randomly picks speakers and multiple utterances per speaker based on the configuration.
• Temporal Arrangement: * In Dialogs, utterances are placed sequentially with configurable pauses.
• In Overlap scenarios, start times are calculated to force speech segments to collide, simulating interruptions.

3. Acoustic Simulation

To improve model generalization, the following augmentations were applied probabilistically:

• Normalization & Fading: RMS-based leveling and fade-ins/outs to prevent digital clipping.
• Volume Variation: Speakers are assigned varied volume levels to simulate different distances from a microphone.
• Reverberation: Room Impulse Responses (RIR) were convolved with the audio to simulate indoor acoustic reflections.
• Background Noise: Additive noise was mixed at Signal-to-Noise Ratios (SNR) ranging from 5dB to 25dB.

Technical Specifications

RTTM Annotation Format

Each audio file is accompanied by an .rttm file following the standard format: SPEAKER <file_id> 1 <start_time> <duration> <NA> <NA> <speaker_id> <NA> <NA>

Suggested Use Cases

• Speaker Diarization: Training EEND or clustering-based models.
• Voice Activity Detection (VAD): Distinguishing speech from silence/noise.
• Overlap Detection: Identifying segments where multiple people speak at once.