nickdee96/ALFFA-Swahili-News

ALFFA Swahili News

Dataset Description

The ALFFA Swahili News dataset is a speech corpus designed for automatic speech recognition (ASR) research in Swahili, an under-resourced African language. This dataset is part of the ALFFA (African Languages in the Field: speech Fundamentals and Automation) project and contains approximately 11.8 hours of broadcast news audio from Radio France International's Swahili service, recorded between November 2010 and March 2011.

The dataset is distributed through OpenSLR (Open Speech and Language Resources) as resource SLR25 and is available from the official OpenSLR repository.

Dataset Summary

This dataset provides:

• Training data: 10,180 audio segments (~10 hours)
• Test data: 1,991 audio segments (~1.8 hours)
• Language: Swahili (Kiswahili)
• Domain: Broadcast news
• Audio format: 16kHz WAV files
• Text: Manual transcriptions in Swahili

Supported Tasks and Leaderboards

• Primary Task: Automatic Speech Recognition (ASR)
• Secondary Tasks:
  • Speaker Identification: Using the speaker_id information for speaker recognition tasks
  • Audio Language Identification: The audio samples can be used as Swahili examples for multi-language identification systems
  • Text-to-Speech (TTS): The paired audio-text data can be used to train TTS models for Swahili speech synthesis
  • Speech-to-text: General speech-to-text applications
  • Phonetic analysis: Linguistic analysis of Swahili pronunciation patterns
  • Voice cloning: Speaker-specific TTS with speaker_id information
  • Speech quality assessment: Using the audio_quality labels for quality detection models
• Evaluation Metrics:
  • ASR: Word Error Rate (WER), Character Error Rate (CER)
  • TTS: MOS (Mean Opinion Score), speaker similarity metrics
  • Language ID: Classification accuracy, F1-score

Languages

The dataset is monolingual, containing only Swahili (ISO 639-1: sw).

Dataset Structure

Data Instances

Each data instance contains:

{
    'id': 'SWH-05-20101106_16k-emission_swahili_05h30_-_06h00_tu_20101106_part10',
    'audio_path': {
        'path': '/path/to/audio.wav',
        'array': array([...], dtype=float32),
        'sampling_rate': 16000
    },
    'text': 'rais wa tanzania jakaya mrisho kikwete',
    'speaker_id': 'SWH-05-20101106', 
    'recording_date': '2010-11-06',
    'start_time': '05:30',
    'end_time': '06:00',
    'part_number': 10,
    'audio_quality': 'good'
}

Data Fields

• id (string): Unique identifier for each audio segment
• audio_path (Audio): Audio file with automatic loading at 16kHz sampling rate
• text (string): Manual transcription of the audio in Swahili
• speaker_id (string): Identifier for the speaker/broadcast session
• recording_date (string): Date of recording (YYYY-MM-DD format)
• start_time (string): Start time of the broadcast segment (HH:MM)
• end_time (string): End time of the broadcast segment (HH:MM)
• part_number (int64): Sequential part number within the broadcast
• audio_quality (string): Quality assessment of the audio

Audio Quality Categories

• good: Clean audio with minimal background noise
• music_background: Audio with background music
• noisy: Audio with background noise
• overlapping_speech: Audio with overlapping speakers
• very_noisy: Very poor quality audio with significant noise
• unknown: Quality not assessed

Data Splits

Split	Examples	Duration
train	10,180	~10 hours
test	1,991	~1.8 hours

Dataset Creation

Curation Rationale

This dataset was created as part of the ALFFA (African Languages in the Field: speech Fundamentals and Automation) project, which aims to develop speech technologies for under-resourced African languages. The project specifically addresses the lack of publicly available speech corpora for African languages by collecting and distributing high-quality speech datasets for Amharic, Swahili, and Wolof.

The Swahili component focuses on broadcast news domain, which provides formal, well-articulated speech suitable for ASR system development. This domain choice ensures consistent audio quality and standardized pronunciation patterns while covering diverse vocabulary relevant to news and current events.

The dataset is part of OpenSLR resource SLR25 and includes Kaldi recipes that are integrated into the main Kaldi distribution, making it easily accessible for researchers working on African language ASR.

Source Data

Initial Data Collection and Normalization

The audio data was collected from Radio France International's Swahili service broadcasts. The recordings were segmented into utterances and manually transcribed by native Swahili speakers.

Who are the source language producers?

The language producers are professional broadcasters and journalists from Radio France International's Swahili service, primarily native speakers of Swahili from East Africa.

Annotations

Annotation process

• Manual transcription by native Swahili speakers
• Quality assessment and labeling of audio segments
• Segmentation into utterance-level chunks
• Speaker identification and metadata extraction

Who are the annotators?

The annotators were native Swahili speakers with expertise in transcription and linguistic analysis, working under the supervision of researchers at GETALP LIG, Grenoble.

Personal and Sensitive Information

The dataset contains public broadcast content and does not include personal or sensitive information beyond what is publicly available in news broadcasts.

Considerations for Using the Data

Social Impact of Dataset

This dataset contributes to the development of speech technology for Swahili, potentially improving access to digital services for Swahili speakers and supporting language preservation efforts.

Discussion of Biases

• Domain bias: Limited to broadcast news, may not generalize to conversational or other domains
• Dialectal bias: May favor certain Swahili dialects used in formal broadcasting
• Gender bias: Speaker distribution may not be balanced across genders
• Quality bias: Varies significantly in audio quality, which may affect model performance

Other Known Limitations

• Limited domain coverage (broadcast news only)
• Relatively small size compared to modern ASR datasets
• Audio quality varies significantly across recordings
• Some segments contain background music or noise
• For TTS: Limited emotional variation (mostly neutral broadcast tone)
• For Language ID: Single language only - needs to be combined with other languages for multi-lingual identification training

Additional Information

Dataset Curators

• Hadrien Gelas - Data collection
• Laurent Besacier - Dataset preparation and curation
• Tien-Ping Tan - Dataset preparation
• François Pellegrino - Research supervision

Licensing Information

This dataset is released under the MIT License, as specified in the OpenSLR repository (SLR25). The original ALFFA project data is distributed under this permissive license to encourage research and development in African language speech technologies.

Citation Information

If you use this dataset, please cite the original paper:

@InProceedings{gelas:hal-00954048,
  author = {Gelas, Hadrien and Besacier, Laurent and Pellegrino, Francois},
  title = {{D}evelopments of {S}wahili resources for an automatic speech recognition system},
  booktitle = {{SLTU} - {W}orkshop on {S}poken {L}anguage {T}echnologies for {U}nder-{R}esourced {L}anguages},
  year = {2012},
  address = {Cape-Town, South Africa},
  url = {http://hal.inria.fr/hal-00954048}
}

For the broader ALFFA project, you may also reference:

@article{gauthier2016collect,
  author = {Gauthier, Elodie and Besacier, Laurent and Voisin, Sylvie and Melese, Michael and Elingui, Uriel Pascal},
  journal = {LREC},
  title = {Collecting Resources in Sub-Saharan African Languages for Automatic Speech Recognition: a Case Study of Wolof},
  year = {2016}
}

Contributions

Thanks to @nickdee96 for converting this dataset to HuggingFace format and making it easily accessible to the research community.

Additional Resources

• OpenSLR Repository: SLR25 - ALFFA
• ALFFA Project: http://alffa.imag.fr
• GitHub Repository: ALFFA_PUBLIC/ASR/SWAHILI
• Download: Original dataset (1.2GB)

Usage Example

from datasets import load_dataset, Audio

# Load the dataset
dataset = load_dataset("nickdee96/swahili-broadcast-news")

# Access train and test splits
train_data = dataset["train"]
test_data = dataset["test"]

# Enable automatic audio loading
train_data = train_data.cast_column("audio_path", Audio(sampling_rate=16000))

# Access a sample
sample = train_data[0]
print(f"Text: {sample['text']}")
print(f"Audio shape: {sample['audio_path']['array'].shape}")
print(f"Sampling rate: {sample['audio_path']['sampling_rate']}")

For Text-to-Speech (TTS)

# The dataset provides high-quality paired audio-text data ideal for TTS training
# Use the text as input and audio as target for TTS model training

# Filter by audio quality for TTS training
good_quality_samples = train_data.filter(lambda x: x['audio_quality'] == 'good')

# Group by speaker for multi-speaker TTS
from collections import defaultdict
speaker_data = defaultdict(list)
for sample in train_data:
    speaker_data[sample['speaker_id']].append(sample)

For Language Identification

# Use as positive examples for Swahili in language identification
# The dataset provides clean Swahili speech samples with quality labels

# Extract audio features for language ID training
swahili_samples = []
for sample in train_data:
    swahili_samples.append({
        'audio': sample['audio_path']['array'],
        'language': 'sw',  # Swahili language code
        'quality': sample['audio_quality']
    })

Baseline Results

The original paper reports the following Word Error Rate (WER) results:

Model	All Audio	Good Quality Only
Monophone	49.28%	38.5%
Triphone	33.55%	23.2%
Triphone + LDA/MLLT	31.92%	22.3%
Triphone + SAT/FMLLR	31.56%	22.4%
Triphone + MMI/fMMI	30.87%	-
SGMM	27.36%	20.7%

These results provide baselines for comparing new ASR systems on this dataset.

Additional Use Cases

Text-to-Speech (TTS) Applications

• Multi-speaker TTS: Use speaker_id to build speaker-specific voice models
• Quality-aware TTS: Filter by audio_quality for high-quality training data
• Prosody modeling: Broadcast speech provides good examples of natural prosody
• Cross-lingual TTS: Use as target language for multilingual TTS systems

Language Identification Applications

• Swahili detection: Use as positive examples in language identification systems
• African language classification: Combine with other African language datasets
• Quality-based filtering: Use audio_quality labels for robust language ID
• Accent variation: Different speakers provide some accent diversity within Swahili

Additional Research Applications

• Speech quality assessment: Use audio_quality labels for automatic quality detection
• Broadcast speech analysis: Study characteristics of news speech vs. conversational speech
• Low-resource language modeling: Bootstrap language models for under-resourced languages
• Cross-domain adaptation: Study domain adaptation from broadcast to other speech types