---
language:
- twi
license: cc-by-nc-4.0
tags:
- tts
- asr
- speech
- ghanaian
- twi
- omnivoice
- synthetic
size_categories:
- 100K<n<1M
task_categories:
- text-to-speech
- automatic-speech-recognition
---

# Ghana Twi Speech Dataset (TTS + ASR)

79,655 synthetic Twi speech samples (88.8 hours) generated with **OmniVoice**. Designed for both **text-to-speech (TTS)** and **automatic speech recognition (ASR)** research on Twi.

## Intended use

- **TTS / Voice cloning**: Use the audio + text pairs directly. The dataset includes diverse voice profiles for building or fine-tuning Twi TTS systems.
- **ASR**: Use the text transcripts as labels. The audio is natively at 24 kHz — resample to 16 kHz for standard ASR pipelines.

## Statistics

| Metric | Value |
|--------|-------|
| Total samples | 79,655 |
| Total duration | 88.78 hours (319,599 seconds) |
| Sample rate | 24 kHz (native) |
| Language | Twi |
| Unique reference speakers | 3158 |
| TTS model | [k2-fsa/OmniVoice](https://huggingface.co/k2-fsa/OmniVoice) |

## Columns

| Column | Type | Description |
|--------|------|-------------|
| `id` | string | Unique sample ID (zero-padded) |
| `audio` | Audio | 24 kHz mono WAV audio |
| `text` | string | Twi transcription |
| `duration` | float | Duration in seconds |
| `ref_speaker` | string | Reference speaker ID used for voice cloning |

## Data generation

- **Texts**: Sampled from [ghananlpcommunity/pristine-twi-english-parallel-sentences](https://huggingface.co/datasets/ghananlpcommunity/pristine-twi-english-parallel-sentences) (Twi column)
- **Voices**: Voice-cloned from 3158 reference speakers across 4 Twi speech corpora (JW, Bible, UNICEF, Finance)
- **Method**: OmniVoice diffusion TTS with random reference audio selection per sample
- **Seed**: 42 (for reproducibility)

## Loading

```python
from datasets import load_dataset

ds = load_dataset("ghananlpcommunity/twi-tts-asr", split="train")
sample = ds[0]
print(sample["text"])         # Twi transcript
print(sample["duration"])     # seconds
print(sample["ref_speaker"])  # which voice was cloned
# -> plays audio
```

### Resampling to 16 kHz for ASR

```python
import torchaudio

def resample_16k(example):
    wav = example["audio"]["array"]
    sr = example["audio"]["sampling_rate"]
    if sr != 16000:
        wav = torchaudio.functional.resample(wav, sr, 16000)
    example["audio_16k"] = wav
    return example

ds_16k = ds.map(resample_16k)
```