atiyehghm/shemo_diarization_persian_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 synthesis framework to simulate realistic conversational dynamics.

• Language: Persian (Farsi)
• Total Duration: ~50 Hours
• Number of Records: 5,000 audio tracks
• Format: 16kHz Mono WAV with structured sequence annotations

Dataset Summary

Metric	Value
Total Audio Files	5,000
Total Duration	50 Hours
Sample Rate	16,000 Hz
Annotation Format	Structured Columns (Start, End, Speaker)
Primary Language	Persian

Conversation Types

The dataset is categorized into three distinct interaction styles:

• Dialog (2,459 tracks): Multi-speaker turn-taking with natural pauses.
• Overlap (2,035 tracks): Competitive speech where multiple speakers talk simultaneously.
• Monologue (506 tracks): Extended speech from a single speaker.

Speaker Distribution

• 1 Speaker: 506 tracks
• 2 Speakers: 883 tracks
• 3 Speakers: 875 tracks
• 4 Speakers: 934 tracks
• 5 Speakers: 906 tracks
• 6 Speakers: 896 tracks

Generation Methodology

1. Source Data

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

2. Composition Algorithm

• 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.
• Acoustic Simulation: Includes RMS-based normalization, volume variation (simulating mic distance), Room Impulse Response (RIR) convolution for reverberation, and additive background noise (SNR 5dB to 25dB).

Technical Specifications

Dataset Structure

Unlike raw RTTM files, this dataset is provided in a structured format compatible with the Hugging Face datasets library. Each record contains the audio data and its corresponding temporal segments:

Feature	Type	Description
audio	audio	Dictionary containing the path and array of the 16kHz audio.
timestamps_start	list[float]	A list of start times (seconds) for each speech segment.
timestamps_end	list[float]	A list of end times (seconds) for each speech segment.
speakers	list[string]	A list of speaker IDs corresponding to each segment.

Annotation Example

A single row in the dataset represents one audio file. The columns contain sequences of events:

• timestamps_start: [0.5, 3.2, 4.1]
• timestamps_end: [2.8, 5.0, 6.5]
• speakers: ["spk_01", "spk_02", "spk_01"]

This format facilitates direct loading into modern diarization pipelines (such as Pyannote or NeMo) without requiring manual parsing of RTTM text files.

Suggested Use Cases

• Speaker Diarization: Training EEND, clustering-based, or modular diarization systems.
• Voice Activity Detection (VAD): Identifying speech vs. non-speech segments.
• Overlap Detection: Training models to identify simultaneous speech segments (where start[i] < end[i-1]).