README.md

---
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: text
    dtype: string
  - name: src_lang
    dtype: string
  - name: tgt_lang
    dtype: string
  - name: duration_seconds
    dtype: float64
  - name: mode
    dtype: string
  - name: speaker
    dtype: string
  splits:
  - name: train
    num_bytes: 0
    num_examples: 8800
  download_size: 0
  dataset_size: 0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: apache-2.0
language:
- en
- fr
- es
- pt
- ar
- sw
- pcm
- yo
- ha
- ig
- ibb
- tiv
- efi
- anw
- idc
- igl
- nup
- ego
- fon
- urh
- tw
- ibr
task_categories:
- automatic-speech-recognition
- text-to-speech
- translation
- audio-classification
tags:
- audio
- speech
- multilingual
- low-resource
- audio-text
- flac
pretty_name: Hypa-Voices
size_categories:
- 1K<n<10K
---

# Hypa-Voices

**Hypa-Voices** is an open-source multilingual, multi-modal audio-text dataset from [Hypa Intelligence](https://huggingface.co/hypaai) and [AfroVoices](https://huggingface.co/hypaai), with a long-term vision of advancing speech and language technology for under-represented languages. Every record pairs written text with corresponding speech, either in the same language (**transcription**) or across two languages while carrying the same meaning (**translation**). In translation records, **speech is in `src_lang` and text is in `tgt_lang`**.

This Hugging Face **collection** releases a curated public subset of the full Hypa-Voices corpus. The full pool contains roughly **2.98 million records**, **8,442 hours**, and **22 source languages**; this release provides **8,800** paired examples spanning those languages so researchers can explore the dataset without downloading the multi-terabyte collection.

**Collection:** [hypaai/hypa-voices](https://huggingface.co/collections/hypaai/hypa-voices)

The collection contains two companion repositories with **aligned metadata** (`text`, `src_lang`, `tgt_lang`, `duration_seconds`, `mode`, `speaker`). They describe the same curated records in two audio representations:

| Repository | Audio format | Best for |
| --- | --- | --- |
| [`hypaai/Hypa-Voices`](https://huggingface.co/datasets/hypaai/Hypa-Voices) (this repo) | **FLAC** (decoded waveform) | Listening, playback, ASR/TTS with raw audio, standard `datasets.Audio` pipelines |
| [`hypaai/Hypa-Voices-snac`](https://huggingface.co/datasets/hypaai/Hypa-Voices-snac) | **SNAC token codes** | Compact storage, token-based speech modeling, Orpheus-style audio-text training |

This dual release supports downstream tasks such as automatic speech recognition (ASR), speech translation, text-to-speech (TTS), cross-lingual transfer, voice cloning, and language modeling over discrete speech tokens.

---

## Dataset Components

### Audio-text pairs (this repository)

- **Source:** Curated subset of the Hypa-Voices corpus, with speech decoded from SNAC back to **FLAC** for direct waveform use.
- **Languages:** Coverage across **22 source languages**, including English, French, Spanish, Portuguese, Arabic, Swahili, Pidgin, and a wide range of under-represented languages such as Ibibio, Tiv, Efik, Yoruba, Hausa, Igbo, Annang, Idoma, Igala, Nupe, Eggon, Fongbe, Urhobo, Twi, and Ebira.
- **Modes:**
  - **`transcribe`:** text and speech share the same language (`src_lang` = `tgt_lang`).
  - **`translate`:** speech is in the source language (`src_lang`) and text expresses the same content in the target language (`tgt_lang`).
- **Format:** Parquet shards with an embedded `audio` feature (FLAC bytes) and shared metadata columns.

### SNAC token companion ([`hypaai/Hypa-Voices-snac`](https://huggingface.co/datasets/hypaai/Hypa-Voices-snac))

- **Source:** Same curated subset and metadata as this repository.
- **Audio representation:** Speech stored as **SNAC** (Multi-Scale Neural Audio Codec) discrete token sequences in the `codes_list` field rather than decoded waveforms.
- **Use case:** Training and evaluation pipelines that operate on compact speech tokens instead of raw audio.

---

## Data Structure

### Data Instances

A typical record contains the transcript, language tags, operational mode, speaker identifier, duration, and audio (FLAC in this repo; SNAC codes in the companion repo).

```python
{
  "audio": {
    "bytes": b"fLaC...",  # FLAC-encoded speech
    "path": ""
  },
  "text": "Onye nlekọta na-enwe nsogbu nọ n'okpuru nrụgide imelite ọzụzụ.",
  "src_lang": "igbo",
  "tgt_lang": "igbo",
  "duration_seconds": 4.9706875,
  "mode": "transcribe",
  "speaker": "SPK-003"
}
```

Translation example (companion SNAC repo uses the same metadata layout):

```python
{
  "codes_list": [130334, 135924, 137439, ...],
  "text": "Mme andiduñ edem usụk ye edem usoputịn Akard edi se ekenọde Hope Medrano Elementary School.",  # tgt_lang (efik)
  "src_lang": "english",   # speech language
  "tgt_lang": "efik",      # text language
  "duration_seconds": 5.2,
  "mode": "translate",
  "speaker": "SPK-111"
}
```

Speaker identifiers in the public release are **anonymised** (`SPK-001`, `SPK-002`, …). The same ID refers to the same voice across records, but contributor names are not included.

### Data Fields

- **audio** (`datasets.Audio`): Decoded speech stored as **FLAC**. When loaded with the `datasets` library, provides `bytes`, `path`, and a decodable waveform array.
- **text** (string): Transcript or translation text paired with the speech clip. For `translate` records, this is the **target-language** side (`tgt_lang`).
- **src_lang** (string): Language of the **speech** in the record.
- **tgt_lang** (string): Language of the **text** in the record (same as `src_lang` for `transcribe` records).
- **duration_seconds** (float): Wall-clock duration of the paired speech in seconds.
- **mode** (string): Either `transcribe` (same-language) or `translate` (cross-language).
- **speaker** (string): Anonymised speaker label (`SPK-XXX`) associated with the speech in the record. The same ID denotes the same voice across rows; contributor names are not published.

### Directory Structure

```
Hypa-Voices/
├── README.md
├── LICENSE
└── data/
    ├── train-00000-of-00022.parquet
    ├── train-00001-of-00022.parquet
    └── ...
```

The SNAC companion follows the same layout, with `codes_list` replacing `audio`:

```
Hypa-Voices-snac/
├── README.md
├── LICENSE
└── data/
    └── train-*.parquet
```

---

## Usage

### Loading with Hugging Face Datasets

```python
from datasets import load_dataset

# Decoded FLAC audio (this repository)
flac_ds = load_dataset("hypaai/Hypa-Voices", split="train")
print(flac_ds[0]["text"], flac_ds[0]["src_lang"], flac_ds[0]["mode"])

# Listen to a clip
flac_ds[0]["audio"]

# SNAC token companion (same metadata columns)
snac_ds = load_dataset("hypaai/Hypa-Voices-snac", split="train")
print(snac_ds[0]["text"], len(snac_ds[0]["codes_list"]))
```

### Loading both repositories together

```python
from datasets import load_dataset

flac_ds = load_dataset("hypaai/Hypa-Voices", split="train")
snac_ds = load_dataset("hypaai/Hypa-Voices-snac", split="train")

# Shared metadata columns
shared_cols = ["text", "src_lang", "tgt_lang", "duration_seconds", "mode", "speaker"]
assert flac_ds.column_names[:1] == ["audio"]
assert snac_ds.column_names[:1] == ["codes_list"]
```

---

## Data Preparation

- **Collection:** Structured AfroVoices recordings, pivot-based translation anchored in English and French, and supplementary open speech material (including Mozilla Common Voice-derived subsets for some languages).
- **Encoding:** Full-corpus speech is represented internally as **SNAC** tokens for efficient storage at scale.
- **Public subset:** A stratified curated sample was selected to cover all **22 source languages** represented in the full corpus.
- **Dual release:** The SNAC repository preserves compact token sequences for modeling; speech in this repository was **decoded from SNAC back to FLAC** so users can work directly with waveforms while keeping identical text and metadata columns.
- **Alignment:** Each row pairs one speech clip with one text transcript. For `translate` records, speech is tagged with `src_lang` and text with `tgt_lang`.
- **Privacy:** Speaker names were replaced with stable anonymised IDs (`SPK-XXX`) before publication. Language associations per speaker are retained for analysis, but contributor identities are not included in the public release.

For a full statistical analysis of the corpus (coverage, speaker concentration, cross-lingual connectivity, transcript structure, and operational balance), see the Hypa-Voices data analysis documentation published alongside this release.

---

## Applications

Hypa-Voices can be used for research and development including, but not limited to:

- **Automatic speech recognition (ASR):** Same-language transcription in under-represented languages.
- **Speech translation:** Cross-lingual pairs with speech in `src_lang` and text in `tgt_lang`, including English/French pivot routes.
- **Text-to-speech (TTS) and voice modeling:** Paired text and speech with speaker labels.
- **Token-based speech modeling:** SNAC-code training via the companion repository.
- **Cross-lingual transfer:** Shared pivot text linking many language pairings.
- **Voice cloning and speaker analysis:** Speaker overlap across languages in the full corpus.

---

## Licensing and Citation

This dataset is released under the **Apache 2.0** license. Please refer to the `LICENSE` file for full details.

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

```bibtex
@misc{hypavoices2026,
  title={Hypa-Voices: Multilingual Low-Resource Audio-Text Dataset},
  author={Hypa Intelligence and AfroVoices},
  year={2026},
  howpublished={Hugging Face collection},
  url={https://huggingface.co/collections/hypaai/hypa-voices}
}
```

If you use the SNAC-encoded companion, please also cite the SNAC codec:

```bibtex
@inproceedings{siuzdak2024snac,
  title={SNAC: Multi-Scale Neural Audio Codec},
  author={Siuzdak, Hubert and Gr{\"o}tschla, Florian and Lanzend{\"o}rfer, Luca A.},
  booktitle={Audio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation},
  year={2024}
}
```

---

## Acknowledgements

- **Hubert Siuzdak** and collaborators, for **SNAC** and for open-sourcing the codec and model weights.
- **Mozilla** and the **Common Voice** community, for open speech data used in parts of the corpus.
- **AfroVoices**, for the steadfast curation of high-quality recordings across under-represented languages.
- **Hypa Intelligence Research (HaIR)**, for our commitment to breaking the language barrier.

---

## Contact and Contributions

For questions, issues, or contributions, please open an issue in this repository or contact **chris@hypaintelligence.com**. Contributions are welcome.

---

## Closing Remarks

By releasing Hypa-Voices, we hope to give researchers and developers practical access to multilingual speech-text data for under-represented languages and to make the structure of that data transparent from the start.

At **Hypa Intelligence**, we believe that for AI to be truly aligned with humanity, it must understand and represent all of us, not just a select few. The first step toward this goal is solving the challenge of multilingualism by **breaking the language barrier**.

Hypa Intelligence remains steadfast in its mission to accelerate the advent of AGI and ASI and to ensure their benefits are globally distributed.

**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. Their mission goes beyond technological innovation, aiming to celebrate the richness of African linguistic diversity on a global stage.