michsethowusu/twi-trigrams-speech-text-parallel

Twi Trigrams Speech-Text Parallel Dataset

Dataset Description

This dataset contains 166156 parallel speech-text pairs for Twi, a language spoken primarily in Ghana. The dataset consists of audio recordings of trigram segments (3-word sequences) paired with their corresponding text transcriptions, making it suitable for automatic speech recognition (ASR) and text-to-speech (TTS) tasks.

Dataset Summary

• Language: Twi - twi
• Task: Speech Recognition, Text-to-Speech
• Size: 166156 trigram audio segments > 1KB (small/corrupted files filtered out)
• Format: WAV audio files with corresponding trigram text labels
• Segment Type: Primarily trigrams (3-word sequences), with some bigrams and single words as fallbacks
• Modalities: Audio + Text

Supported Tasks

• Automatic Speech Recognition (ASR): Train models to convert Twi speech to text
• Text-to-Speech (TTS): Use parallel data for TTS model development
• Keyword Spotting: Identify specific Twi word sequences in audio
• N-gram Language Modeling: Study Twi trigram patterns
• Phonetic Analysis: Study Twi pronunciation patterns in context

Dataset Structure

Data Fields

• audio: Audio file in WAV format containing a trigram segment
• text: Corresponding text transcription (typically 3 words, sometimes 2 or 1 for shorter segments)

Data Splits

The dataset contains a single training split with 166156 filtered trigram audio segments.

Dataset Creation

Source Data

The audio data has been sourced ethically from consenting contributors. To protect the privacy of the original authors and speakers, specific source information cannot be shared publicly.

Data Processing

1. Audio Alignment: Original audio files were processed using forced alignment to obtain word-level timestamps
2. Trigram Segmentation: Audio was segmented into overlapping trigrams (3-word sequences)
3. Fallback Segmentation: For shorter texts, bigrams or single words were created as needed
4. Quality Filtering:
   • Segments longer than 30 seconds were excluded
   • Segments shorter than 0.1 seconds were excluded
   • Files smaller than 1KB were filtered out to ensure audio quality
5. Text Processing: Text was lowercased and cleaned of end punctuation
6. Unique Naming: Each segment received a unique sequential filename (trigram_XXXXXX.wav)

Alignment Technology

Audio processing and word-level alignment performed using the MMS-300M-1130 Forced Aligner tool, which provides accurate timestamp information for creating precise trigram segments.

Annotations

Text annotations represent the spoken content in each trigram audio segment, with text processing applied for consistency:

• Lowercased for uniformity
• End punctuation removed
• Spaces normalized

Considerations for Using the Data

Social Impact of Dataset

This dataset contributes to the preservation and digital representation of Twi, supporting:

• Language technology development for underrepresented languages
• Educational resources for Twi language learning
• Cultural preservation through digital archives
• N-gram based language modeling research

Discussion of Biases

• The dataset may reflect the pronunciation patterns and dialects of specific regions or speakers
• Audio quality and recording conditions may vary across segments
• Trigram distribution may not be representative of natural Twi language patterns
• Some segments may contain overlapping content due to the sliding window approach

Other Known Limitations

• Segment-level rather than full sentence context
• Potential audio quality variations between segments
• Regional dialect representation may be uneven
• Variable segment lengths (primarily 3 words, but includes 2-word and 1-word segments)

Additional Information

Dataset Statistics

• Primary Content: Trigram segments (3-word sequences)
• Fallback Content: Bigram segments (2-word sequences) and single words
• Segment Duration: 0.1 to 30 seconds
• Minimum File Size: 1KB after processing

Licensing Information

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Citation Information

If you use this dataset in your research, please cite:

@dataset{twi_trigrams_parallel_2025,
  title={Twi Trigrams Speech-Text Parallel Dataset},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/michsethowusu/twi-trigrams-speech-text-parallel}}
}

Acknowledgments

• Audio processing and alignment performed using MMS-300M-1130 Forced Aligner
• Forced alignment and trigram segmentation using CTC forced alignment techniques
• Thanks to all contributors who provided audio samples while maintaining privacy protection

Contact

For questions or concerns about this dataset, please open an issue in the dataset repository.