README.md

---
language: [ak, tw, fat]
license: cc-by-4.0
multilinguality: monolingual
task_categories: [text-to-speech, automatic-speech-recognition]
tags: [akan, twi, fante, speech, tts, asr, african-languages, low-resource]
pretty_name: Akan Speech Data (Pooled)
size_categories: [10K<n<100K]
---

# Akan Speech Data (Pooled)

A **~75.4-hour** Akan speech corpus, pooling a crowdsourced ASR config with two
single-narrator studio TTS configs for Akan's Twi and Fante dialects. Part of the
[AfroNet](https://github.com/osinkolu/afronet-tts-data) multi-language TTS data
effort.

## Source

[**WAXAL**](https://huggingface.co/datasets/google/WaxalNLP) (`google/WaxalNLP`),
three configs pooled together:

- `aka_asr` — crowdsourced, image-prompted speech, many speakers. 12,751 clips, 69.5h.
- `twi_tts` — single-narrator studio-quality read speech, Twi dialect. 793 clips, 3.0h.
- `fat_tts` — single-narrator studio-quality read speech, Fante dialect. 874 clips, 2.9h.

Twi and Fante are Akan dialects (mutually intelligible to a high degree); they're
pooled here as one "Akan" release rather than split into separate languages. The
`domain` field in the manifest preserves which config each clip came from
(`aka_asr`/`twi_tts`/`fat_tts`), so you can filter by dialect/register if needed.

`train`+`validation`+`test` splits are pooled together across all three configs
(intentional, same policy already applied to WAXAL's other `_tts` configs used
elsewhere in AfroNet); the `unlabeled` split that exists for `aka_asr` (untranscribed
audio) is excluded, consistent with AfroNet's no-auto-transcription policy.

All audio is standardized to **16 kHz mono FLAC** (lossless), 1–30 second clips.

## Format

The dataset ships as **WebDataset-style tar shards** (`shards/shard-00000.tar` …, ~1 GB
each, one `{key}.flac` file per clip) plus a single manifest (`manifest.parquet` /
`manifest.jsonl`):

| Column | Description |
|---|---|
| `key`, `shard` | which tar file + entry holds this clip's audio |
| `text` | transcript (native script) |
| `duration` | seconds |
| `source` | always `waxal` |
| `dataset_id` | always `0` |
| `split` | `train` / `val` (250 clips held out for evaluation) |
| `speaker_id` | source-provided speaker ID |
| `gender` | speaker metadata where available |
| `domain` | which WAXAL config the clip came from: `aka_asr`, `twi_tts`, or `fat_tts` |
| `dbfs`, `clip_ratio`, `sil_ratio` | cheap DSP quality proxies: loudness, fraction of clipped samples, fraction of near-silent frames |
| `has_disfluency` | always `false` |

## Usage

```python
from huggingface_hub import hf_hub_download
import pandas as pd, tarfile, io, soundfile as sf

mp = hf_hub_download("Professor/akan-speech-data", "manifest.parquet", repo_type="dataset")
df = pd.read_parquet(mp)

row = df.iloc[0]
shard_path = hf_hub_download("Professor/akan-speech-data", f"shards/{row.shard}", repo_type="dataset")
with tarfile.open(shard_path) as tar:
    audio_bytes = tar.extractfile(f"{row.key}.flac").read()
arr, sr = sf.read(io.BytesIO(audio_bytes))
```

The tar shards are also directly readable by the [`webdataset`](https://github.com/webdataset/webdataset)
library for streaming training pipelines.

## Intended use & limitations

Built for **Akan TTS/ASR research**, in particular as finetuning data for a
multilingual TTS model that doesn't natively support Akan. A small slice of this
corpus (the two `_tts` configs, ~8% of hours) is clean single-speaker studio read
speech; the large majority (`aka_asr`) is crowdsourced, image-prompted speech with
many speakers and more variable recording conditions. This is a **research
aggregation**; usage should respect WAXAL's own terms.

## License

CC BY 4.0, per the upstream [WAXAL](https://huggingface.co/datasets/google/WaxalNLP) release.

## Acknowledgments

Deep thanks to the **WAXAL** project (Google) and its Akan/Twi/Fante-speaking
contributors for the source corpus.

This dataset was pooled by **Victor Olufemi and LyngualLabs** as part of the
[AfroNet](https://github.com/osinkolu/afronet-tts-data) multi-language TTS data effort.