BeitTigreAI/tigre-speech-text-aligned

Tigre Speech Corpus

1. Overview

This Tigre Speech Corpus is a curated collection of 18,470 aligned audio–text pairs designed to support research and development in speech technologies for Tigre (tig), an under-resourced South Semitic language spoken primarily in Eritrea. The dataset contains approximately 32 hours of recorded speech contributed by over 100 native speakers.

It reflects a collective effort by Tigre-speaking contributors worldwide, including a significant number of diaspora native speakers who participated in the Mozilla Common Voice project. This corpus is intended to serve as a foundational resource for advancing NLP technologies for the Tigre language and is suitable for tasks such as:

• Automatic Speech Recognition (ASR)
• Forced alignment
• Speech translation
• Speech-to-text pretraining
• Low-resource speech benchmark creation

---

2. Data Source, Provenance, and Licensing

The data was voluntarily collected via the Mozilla Common Voice platform, with contributions from community members using personal devices (mobile phones and computers). The sentence prompts used for recording originate from the Common Voice sentence bank.

Licensing

The dataset follows Mozilla Common Voice licensing terms:

• Audio recordings (WAV files): CC0 1.0 Universal — Public Domain Dedication
• Text content: CC BY 4.0 — Attribution required

---

3. Corpus Composition

3.1 Data Pairing Structure

Each item in the corpus consists of:

• A .wav file (audio recording)
• A .txt file (transcript)

Both share the same filename prefix (e.g., clip_00123.wav / clip_00123.txt), ensuring straightforward alignment.

3.2 Duration Distribution

The speech duration distribution:

• ~⅓ of audio clips: < 5 seconds
• ~⅓ of audio clips: 5–7 seconds
• ~⅓ of audio clips: 7–41 seconds

This natural variation supports both short- and long-utterance modeling.

---

4. Dataset Structure and File Format

All files are stored in a flat directory structure inside the compressed release file.
Each audio file has a single matching transcript with the exact same filename prefix.

File Layout

dataset_root/
├── clip_xxxxx.wav
├── clip_xxxxx.txt
├── clip_yyyyy.wav
├── clip_yyyyy.txt
└── ...

• .wav files contain audio recordings.
• .txt files contain transcripts.
• There are no nested folders to simplify ingestion by ASR pipelines.

Naming Convention

Filename alignment pattern:

• clip_12345.wav
• clip_12345.txt

This ensures 1:1 audio–text alignment for all 18,470 entries.

---

5. Intended Use & Applications

This dataset supports:

• Low-resource ASR development
• End-to-end speech models (wav2vec2, Whisper, MMS)
• African language benchmarks
• Cross-lingual representation learning
• Phonetic/linguistic analysis

Potential applications:

• Voice assistants
• Speech-enabled educational tools
• Accessibility technologies
• Humanitarian language technology

---

6. Ethical Considerations

• All recordings were contributed voluntarily.
• Contributors acknowledged CC0 audio licensing.
• No personally identifiable information (PII) is included.
• Derivative models should respect the spirit of open, community-driven data.

---

7. Recommended Citation

@dataset{tigre_speech_corpus_2025,
  title = {Tigre Speech Corpus},
  author = {Tigre Diaspora Community Contributors and Mozilla Common Voice},
  year = {2025},
  url = {https://huggingface.co/},
  note = {A collection of 18,470 audio–text pairs for the Tigre language.}
}

---

8. Acknowledgments

We gratefully acknowledge:

• The Tigre diaspora community for their contributions.
• The Mozilla Common Voice team for enabling community-driven speech data creation.