Aybee5/ha-tts-mixed

Dataset conversion helper

1. Install dependencies:

   pip install -r requirements.txt

2. Run the script from the repo root to create the parquet and copy audio files into data/audio_files:

   python3 scripts/create_parquet.py

3. Resulting files:

   • data/dataset.parquet (contains columns: source, text, audio)
   • data/audio_files//.wav (copied audio files)

Notes:

• The audio column contains relative paths starting with data/audio_files/... so the whole data/ folder can be uploaded to Hugging Face or copied into Colab and loaded with relative paths.
• If you want to only inspect without copying, run with --dry-run.

Dataset Card

This dataset is designed for text-to-speech (TTS) applications. It contains audio files along with their corresponding text.

Example for Audio Rendering

To render audio from the dataset, you can use the following code snippet:

import soundfile as sf
import numpy as np

# Load audio file
audio_path = 'data/audio_files/<speaker_id>/<audio_id>.wav'
audio_data, sample_rate = sf.read(audio_path)

# Play audio (requires sounddevice library)
import sounddevice as sd
sd.play(audio_data, sample_rate)
sd.wait()  # Wait until audio is finished playing

---

dataset: name: "ha-tts-mixed" description: | Hausa mixed single/multi-speaker TTS dataset converted from local MimicStudio export. Each row has the speaker id, the text transcript and a relative path to the audio file. license: "unspecified" features: - name: source type: string description: Speaker id (used as source in Unsloth format) - name: text type: string description: Transcript / prompt text for the audio file - name: audio type: audio description: Relative path to the wav file under data/audio_files/...

How the Hugging Face dataset card will render

On the Hugging Face dataset page the audio column will be playable inline, and the text and source fields will be shown alongside it. Example row:

source	text	audio
c3621689-ca53-c1e1-d0c1-e5619d6c0634	Ansamu ɓaraka acikin shirin	(playable audio)

If you'd like, I can add a full dataset_card.md in the repo root with licensing, citation, and a more detailed schema.