GeoPoll/dataset-20250728_102101-sw

GeoPoll Swahili Speech Dataset

This dataset contains speech recognition data for Swahili (sw) collected and processed by GeoPoll.

Dataset Summary

This dataset is designed for fine-tuning speech recognition models on Swahili audio data. It includes high-quality audio segments with corresponding transcriptions.

Dataset Statistics

• Total samples: 11814
• Total duration: 20.45 hours
• Average duration: 6.23 seconds per sample
• Number of speakers: 6
• Language: Swahili (sw)
• Quality score: 0.751/1.0

Data Structure

Each sample contains:

• audio: Audio file path and metadata
• text: Transcribed text
• speaker: Speaker identifier
• metadata: Additional information including duration, timestamps, etc.

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("GeoPoll/dataset-20250728_102101-sw")

Fine-tuning Whisper

from transformers import WhisperProcessor, WhisperForConditionalGeneration
from datasets import load_dataset

# Load dataset
dataset = load_dataset("GeoPoll/dataset-20250728_102101-sw")

# Load model and processor
processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

# Preprocess function
def preprocess_function(examples):
    audio_arrays = [x["array"] for x in examples["audio"]]
    inputs = processor(
        audio_arrays, 
        sampling_rate=16000, 
        return_tensors="pt", 
        padding=True
    )
    
    labels = processor.tokenizer(
        examples["text"], 
        return_tensors="pt", 
        padding=True
    ).input_ids
    
    return {
        "input_features": inputs.input_features,
        "labels": labels
    }

# Apply preprocessing
dataset = dataset.map(preprocess_function, batched=True)

Fine-tuning Wav2Vec2

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset

# Load dataset
dataset = load_dataset("GeoPoll/dataset-20250728_102101-sw")

# Load model and processor
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base")

# Preprocess function
def preprocess_function(examples):
    audio_arrays = [x["array"] for x in examples["audio"]]
    inputs = processor(
        audio_arrays, 
        sampling_rate=16000, 
        return_tensors="pt", 
        padding=True
    )
    
    labels = processor.tokenizer(
        examples["text"], 
        return_tensors="pt", 
        padding=True
    ).input_ids
    
    return {
        "input_values": inputs.input_values,
        "labels": labels
    }

# Apply preprocessing
dataset = dataset.map(preprocess_function, batched=True)

Data Collection

This dataset was collected by GeoPoll through:

• Voice calls and surveys
• High-quality audio recording
• Professional transcription services
• Speaker diarization and segmentation

Data Processing

The audio data has been processed with:

• Noise reduction and normalization
• Segmentation based on speaker turns
• Text cleaning and formatting
• Quality filtering (minimum duration, word count, etc.)

Quality Control

• Minimum segment duration: 2 seconds
• Minimum words per segment: 4
• Manual review of transcriptions
• Automatic quality scoring

Licensing

Copyright (c) 2025 GeoPoll. All rights reserved.

Citation

If you use this dataset, please cite:

@dataset{geopoll_sw_speech_dataset,
  title={GeoPoll Swahili Speech Dataset},
  author={GeoPoll},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/GeoPoll/dataset-20250728_102101-sw}
}

Contact

For questions or issues with this dataset, please contact GeoPoll at support@geopoll.com

---

This dataset was created using GeoPoll's speech data collection and processing pipeline.