Hypa-Voices-ASR-Benchmark

AfroVoices, a subsidiary of Hypa AI, is dedicated to amplifying African voices, languages, and cultures in the intelligence age. Focused on bridging the digital representation gap, AfroVoices curates datasets and resources for African languages, promoting inclusivity and cultural appreciation in AI technologies.

Dataset Summary

Hypa-Voices-ASR-Benchmark is a compact, 20-sample benchmark drawn from the larger hypaai/Hypa-Voices corpus. It is designed as a lightweight, quick-to-run evaluation set for automatic speech recognition (ASR) systems, covering a mix of major world languages and low-resource African languages. Each sample pairs an audio recording with its verbatim transcription, making it well suited for rapid sanity checks, smoke tests, and small-scale ASR benchmarking before scaling up to the full parent dataset.

Source Data

All samples are a curated subset of the Hypa-Voices dataset, which was built by AfroVoices through professional transcription and recording of audio-text pairs spanning English, French, Spanish, and a range of low-resource African languages including Yoruba, Igbo, Hausa, Tiv, Ibibio, Annang, Efik, and others. The 20 examples in this benchmark were selected to preserve the language diversity of the parent dataset while keeping the download and evaluation footprint minimal.

Dataset Structure

Data Fields

• audio (datasets.Audio): The speech recording to be transcribed (FLAC format).
• text (string): The reference transcription of the audio.
• src_lang (string): The language of the spoken audio.
• tgt_lang (string): The language of the reference transcription (matches src_lang for transcription samples).
• duration_seconds (float64): Duration of the audio clip, in seconds.

Note: this benchmark intentionally omits the mode and speaker fields present in the full Hypa-Voices dataset, to keep the card focused purely on ASR evaluation.

Data Splits

Split	Samples
train	20

Use Cases

This dataset can be used for:

• Automatic Speech Recognition (ASR): Quick evaluation of transcription models (e.g., Word Error Rate benchmarking) across a mix of high- and low-resource languages.
• Regression / Smoke Testing: A small, fast-loading benchmark for CI pipelines or early-stage model checks before scaling to the full Hypa-Voices dataset.
• Qualitative Review: Spot-checking transcription quality and error patterns across diverse African and world languages.

Limitations

Given its small size (20 samples), this benchmark is not intended to serve as a statistically robust evaluation set. It is best used for quick iteration and sanity checks; for rigorous benchmarking, please use the full hypaai/Hypa-Voices dataset.

License

This dataset is released under the Apache 2.0 License.

Citation

When using Hypa-Voices-ASR-Benchmark in your work, please cite:

@misc{hypavoicesasrbenchmark,
  title={Hypa-Voices-ASR-Benchmark: A Mini Multilingual Speech Recognition Benchmark for African and World Languages},
  author={Hypa AI and AfroVoices},
  year={2025},
  note={Open-sourced on Hugging Face},
  url={https://huggingface.co/datasets/hypaai/Hypa-Voices-ASR-Benchmark}
}

Please also acknowledge the parent dataset:

@misc{hypavoices,
  title={Hypa-Voices: A Multilingual Speech and Text Dataset for African and World Languages},
  author={Hypa AI and AfroVoices},
  year={2025},
  note={Open-sourced on Hugging Face},
  url={https://huggingface.co/datasets/hypaai/Hypa-Voices}
}

Contact

For any questions, issues, or contributions, please open an issue in this repository or contact hypa.ai.ng@gmail.com. Contributions are welcome!