psdn-ai/psdn-voice-samples-urdu

Poseidon Urdu Speech Dataset

Dataset Description

This dataset contains 2,500 Urdu audio recordings from the Poseidon audio campaign.

Dataset Statistics

• Total Samples: 2,500
• Total Duration: 27.98 hours
• Average WER: 0.3335
• Average CER: 0.2222
• Average Semantic Score: 0.9324
• Average Poseidon Score: 0.8084

Decision Rules

• Poseidon score (poseidon_score) > 0.65 (higher the better)

Language Distribution

• ur: 2,500 samples

Dataset Structure

Data Fields

• audio: Audio file metadata and bytes
• file_id: Unique identifier for the audio file
• speaker_id: Unique identifier for the speaker
• language_code: ISO language code
• GT_transcript_native: Ground truth transcript in Urdu
• GT_transcript_english: Ground truth transcript in English
• spoken_transcript_native: ASR-generated transcript in Urdu
• spoken_transcript_english: ASR-generated transcript translated to English
• wer_score: Word Error Rate score (range: [0,1])
• cer_score: Character Error Rate score (range: [0,1])
• semantic_score: Semantic similarity score (range: [0,1])
• poseidon_score: Overall quality score (range: [0,1])
• duration: Audio duration in seconds
• sampling_rate: Audio sampling rate in Hz
• embedding: a 192 dimensional embedding generated from https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb for the audio file
• jitter: Jitter (local absolute) percentage - voice quality metric measuring pitch period variation (range: 0.66-4.32%, mean: 1.70%)

Data Splits

The dataset is delivered as a single train split (100% of the data).

Usage

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("psdn-ai/psdn-voice-samples-urdu")

# Load specific split
train_data = load_dataset("psdn-ai/psdn-voice-samples-urdu", split="train")

# Access audio and metadata
sample = dataset["train"][0]
audio_array = sample["audio"]["array"]
sampling_rate = sample["audio"]["sampling_rate"]
transcript = sample["GT_transcript"]
duration = sample["duration"]

Quality Metrics

This dataset bundles multiple quality indicators:

• WER (Word Error Rate): Measures word-level transcription accuracy
• CER (Character Error Rate): Measures character-level transcription accuracy
• Semantic Score: Measures semantic similarity between spoken and reference transcripts
• Poseidon Score: Composite quality score derived from the above metrics

Filtering Examples

from datasets import load_dataset

dataset = load_dataset("psdn-ai/psdn-voice-samples-urdu", split="train")

# Filter clips with low spam probability
human_sounding = dataset.filter(lambda x: x["poseidon_score"] > 0.65)

Citation

@dataset{poseidon_urdu_speech_dataset_2025,
  title={Poseidon Urdu Speech Dataset},
  author={Poseidon-AI},
  year={2025},
  publisher={Poseidon-AI},
  howpublished={\url{https://huggingface.co/datasets/psdn-ai/psdn-voice-samples-urdu}}
}

Contact

For questions or issues, please contact Poseidon team.