repleeka/Tagin-Speech-Corpus

---

Tagin Speech Corpus (Taliha Circle)

Dataset Description

The Tagin Speech Corpus is a raw, uncurated collection of speech recordings in the Tagin language (ISO 639-3: tgj). The data was primarily sourced from native speakers conversing in WhatsApp groups belonging to the Taliha Circle region in the Upper Subansiri district of Arunachal Pradesh, India.

This dataset is intended to support the development and research of Automatic Speech Recognition (ASR) systems and linguistic analysis for the Tagin language, a low-resource language belonging to the Tani branch of the Sino-Tibetan language family.

Dataset Details

• Language: Tagin
• ISO 639-3 Code: tgj
• Geographic Area: Taliha Circle, Upper Subansiri district, Arunachal Pradesh, India
• Source: Crowdsourced / WhatsApp Group audio recordings
• Total Duration: 3 hours, 48 minutes, and 25 seconds
• Total Files: 267 audio files
• Data Type: Audio (.wav) accompanied by text transcriptions (metadata.csv)

Audio Profile

The audio files have been converted to a standardized format, but retain their original acoustic environments and noise profiles.

• Format: .wav
• Sampling Rate: 16,000 Hz (16 kHz)
• Channels: 1 (Mono)
• Bit Depth: 16-bit PCM

Dataset Structure

The corpus is organized as follows:

├── audio/
│   ├── Taliha_speech001.wav
│   ├── Taliha_speech002.wav
│   └── ...
└── metadata.csv

metadata.csv

The metadata file maps each audio file to its corresponding transcription. It contains the following columns:

• file_name: The relative path to the audio file (e.g., audio/Taliha_speech001.wav).
• transcription: The text transcription of the spoken audio.

Important Usage Note: Preprocessing Required

This audio is currently unprocessed and is not yet suitable for direct ASR model training.

Because the data originates from raw WhatsApp voice notes, it contains real-world artifacts such as:

• Background noise and varying acoustic environments
• Potential clipping or varying volume levels
• Conversational disfluencies, pauses, and overlapping background speech

Researchers and developers will need to perform extensive data cleaning and preprocessing before utilizing this dataset for training downstream machine learning models. Recommended preprocessing steps include:

• Voice Activity Detection (VAD) to trim long silences
• Noise reduction and audio normalization
• Alignment verification between the transcriptions and the audio

Social Impact and Ethics

This dataset contributes to the preservation and technological inclusion of the Tagin language. By open-sourcing these linguistic resources, we hope to encourage NLP and Speech researchers to build tools for underrepresented and low-resource communities.

Dataset Curators

Names: Tungon Dugi and Jurist Dupit

Institute: National Institute of Technology Arunachal Pradesh
Department: Computer Science and Engineering
Contact: tungondugi@gmail.com