README.md

---
language:
- plt
license: cc-by-nc-sa-4.0
pretty_name: Mimba StyleTTS2 PLT Corpus (Plateau Malagasy TTS Training Corpus)
task_categories:
- text-to-speech
tags:
- text-to-speech
- tts
- Plateau Malagasy
- plt
- malagasy
- african-languages
- low-resource
- StyleTTS2
- end-to-end-tts
size_categories:
- 10K<n<100K
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: sample_rate
      dtype: int64
    - name: text_original
      dtype: string
    - name: text_phonemized
      dtype: string
    - name: speaker_id
      dtype: int64
    - name: speaker_name
      dtype: string
configs:
- config_name: default
  data_files:
  - split: train
    path: chunk_*.parquet
---

# Mimba StyleTTS2 PLT Corpus — Plateau Malagasy TTS Training Corpus

A **ready-to-train corpus for StyleTTS2** on **Plateau Malagasy (PLT)**, pairing
clean audio, original text, and IPA-phonemized text for one or more reference
speakers. Built as a unified, self-contained HuggingFace dataset so that training
notebooks can load a single source of truth — audio, transcription and speaker
metadata in one place — without juggling multiple files or repositories.

> ⚠️ **Derived from synthetic audio.** The audio in this corpus originates from
> [`mimba/plt-tts-dataset`](https://huggingface.co/datasets/mimba/plt-tts-dataset),
> which was generated by [OmniVoice](https://github.com/k2-fsa/OmniVoice). It is
> **not** human-recorded speech. See [Limitations](#limitations-and-known-issues).

---

## Dataset summary

| Property | Value |
|---|---|
| Language | Plateau Malagasy (Merina) / `plt` |
| Task | Text-to-speech (TTS) training — StyleTTS2 Stage 1 & 2 |
| Audio | Mono, **24 000 Hz**, PCM-16 WAV bytes embedded in Parquet |
| Speakers | Extensible — currently `spk_m2` (id=0), `spk_f1` (id=2) |
| Samples | ~38 700 per speaker (derived from `mimba/plt-tts-dataset`) |
| Phonemization | IPA via **tafitaribhi** (PLT G2P), mode PHRASE, normalized with `normalize_plt` |
| Source audio | [`mimba/plt-tts-dataset`](https://huggingface.co/datasets/mimba/plt-tts-dataset) |
| Source text | Same as above — ultimately from [`mimba/text2text`](https://huggingface.co/datasets/mimba/text2text) (`plt_fra`) |

---

## ⚡ How to use

The `audio` column is stored as `{"bytes": <WAV file bytes>, "path": None}` —
identical to the standard 🤗 `datasets` `Audio` feature layout.

### Option A — Let `datasets` decode it (recommended)

```python
from datasets import load_dataset, Audio

ds = load_dataset("mimba/styletts2-plt-corpus", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=24000))

sample = ds[0]
print(sample["speaker_name"], "|", sample["speaker_id"])
print("original  :", sample["text_original"])
print("phonemized:", sample["text_phonemized"])

audio = sample["audio"]   # {'array': np.float32[...], 'sampling_rate': 24000, 'path': None}
print(audio["array"].shape, audio["sampling_rate"])
```

### Option B — Decode the bytes yourself

```python
import io, soundfile as sf
from datasets import load_dataset

ds = load_dataset("mimba/styletts2-plt-corpus", split="train")
sample = ds[0]

array, sr = sf.read(io.BytesIO(sample["audio"]["bytes"]))
print(array.shape, sr)   # e.g. (57600,) 24000
```

### Play it in a notebook

```python
import IPython.display as ipd
ipd.display(ipd.Audio(array, rate=sr))
```

### Filter by speaker

```python
ds = load_dataset("mimba/styletts2-plt-corpus", split="train")

spk_m2 = ds.filter(lambda x: x["speaker_name"] == "spk_m2")
spk_f1 = ds.filter(lambda x: x["speaker_name"] == "spk_f1")
```

### Build `train_list.txt` / `val_list.txt` for StyleTTS2

StyleTTS2's `meldataset.py` expects `wav_path|phonemes|speaker_id` filelists.
This corpus is designed to produce them with a single iteration, without any
extra phonemization step:

```python
import io, os, soundfile as sf
from datasets import load_dataset

ds = load_dataset("mimba/styletts2-plt-corpus", split="train")
os.makedirs("wavs", exist_ok=True)
lines = []

for i, sample in enumerate(ds):
    uid = f"{sample['speaker_name']}_{i:06d}.wav"
    with open(f"wavs/{uid}", "wb") as f:
        f.write(sample["audio"]["bytes"])
    lines.append(f"{uid}|{sample['text_phonemized']}|{sample['speaker_id']}")

# split 95/5 train/val
n_val = max(4, len(lines) // 20)
open("Data/train_list.txt", "w").write("\n".join(lines[:-n_val]))
open("Data/val_list.txt",   "w").write("\n".join(lines[-n_val:]))
```

### Stream (large corpus)

```python
from datasets import load_dataset

ds = load_dataset("mimba/styletts2-plt-corpus", split="train", streaming=True)
for sample in ds:
    print(sample["text_phonemized"])
    break
```

---

## Data fields

| Field | Type | Description |
|---|---|---|
| `audio` | `dict` | `{"bytes": <WAV bytes, 24 kHz mono, PCM-16>, "path": None}` |
| `sample_rate` | `int` | Always `24000` |
| `text_original` | `string` | Original PLT sentence before normalization |
| `text_phonemized` | `string` | IPA phonemization (tafitaribhi, mode PHRASE, post-normalized) |
| `speaker_id` | `int` | Integer speaker index (0 = `spk_f1`, 1 = `spk_f2`, 2 = `spk_m1`, 3 = `spk_m2`, …) |
| `speaker_name` | `string` | Human-readable speaker id (e.g. `spk_m2`) |

Chunk files are named **`chunk_{speaker_name}_{index:05d}.parquet`** — one
speaker's data never overwrites another's, making it safe to add new speakers
to the same repository without any migration.

---

## Speakers

| `speaker_name` | `speaker_id` | Gender (label) | Source |
|---|---|---|---|
| `spk_m2` | 2 | M | `mimba/plt-tts-dataset` (spk_m2 shards) |
| `spk_f1` | 0 | F | `mimba/plt-tts-dataset` (spk_f1 shards) |

> ⚠️ **Gender labels are inherited** from the source dataset and were not
> re-verified acoustically. Confirm by listening if gender is relevant to
> your use case.

---

## How the data was produced

1. **Source audio** — the raw waveforms come from
   [`mimba/plt-tts-dataset`](https://huggingface.co/datasets/mimba/plt-tts-dataset),
   itself generated by [OmniVoice](https://github.com/k2-fsa/OmniVoice) in
   clone-by-reference mode (`language_id='plt'`). Only the target speakers
   are included here.

2. **Audio processing** — each waveform is resampled to 24 000 Hz if needed,
   converted to float32 mono, peak-normalized to 0.95, and re-encoded as
   PCM-16 WAV bytes. Entries with zero-length audio (corrupted source) are
   silently dropped.

3. **Text normalization** — `normalize_plt` converts numbers to PLT words
   (e.g. `25 → dimy amby roapolo`), expands common abbreviations, and
   normalizes punctuation. Digits embedded in usernames or hashtags (e.g.
   `Zaw2`, `#100Africanmyths`) are left untouched by a word-boundary guard
   (`\b\d+\b`) to avoid producing nonsense words.

4. **Phonemization** — normalized text is converted to IPA by
   **tafitaribhi** (a PLT-specific G2P engine), called in **phrase mode**
   (full sentence at once, preserving cross-word prosody). Curly apostrophes
   (U+2019) produced by the phonemizer are normalized to straight apostrophes
   to match the symbol vocabulary used by PL-BERT and StyleTTS2. Characters
   outside the 55-symbol IPA vocabulary are stripped.

5. **Storage** — examples are written in Parquet shards of 1 000 rows each,
   scoped by speaker (`chunk_{speaker_name}_{index:05d}.parquet`), and pushed
   incrementally to this repository. The `OOD_texts.txt` file (50 000 IPA
   sentences from `mimba/pl-bert-phonemized-plt`) is also included for use as
   StyleTTS2's out-of-distribution text reference.

---

## Limitations and known issues

- **Synthetic, not human.** Audio inherits OmniVoice's PLT pronunciation and
  prosody artifacts; it is not a substitute for real recorded speech.
- **Source-text bias.** The underlying text is predominantly biblical/religious
  in register; everyday conversational vocabulary is under-represented.
- **Phonemization is approximate.** tafitaribhi models PLT G2P rules and may
  make errors on rare words, loanwords, or numerals not caught by `normalize_plt`.
- **No per-sample human validation.** Spot-check alignment between audio and
  `text_phonemized` before training at scale.
- **IPA vocabulary fixed at 55 symbols.** Characters outside this set are
  stripped; very rare IPA symbols may be lost.

---

## Relation to other Mimba datasets

| Dataset | Role |
|---|---|
| [`mimba/text2text`](https://huggingface.co/datasets/mimba/text2text) | Original PLT source text |
| [`mimba/plt-tts-dataset`](https://huggingface.co/datasets/mimba/plt-tts-dataset) | Raw multi-speaker synthetic audio (4 speakers, full corpus) |
| [`mimba/pl-bert`](https://huggingface.co/datasets/mimba/pl-bert) | 3.9M PLT sentences for PL-BERT pre-training |
| [`mimba/pl-bert-phonemized-plt`](https://huggingface.co/datasets/mimba/pl-bert-phonemized-plt) | PL-BERT corpus phonemized in phrase mode |
| [`mimba/pl-bert-plt-final`](https://huggingface.co/datasets/mimba/pl-bert-plt-final) | PL-BERT training-ready dataset + vocabularies (`phoneme_symbols.pkl`, `word_vocab.pkl`) |
| **`mimba/styletts2-plt-corpus`** ← *this dataset* | StyleTTS2-ready corpus (audio + IPA + speaker_id) |

---

## Intended use

This dataset was built to train **StyleTTS2** (Stage 1 and Stage 2) on
Plateau Malagasy, as part of the **Mimba** offline TTS project for low-resource
African languages. The `text_phonemized` column is directly consumable by
StyleTTS2's `meldataset.py` without any additional phonemization step.

It can also serve as a fine-tuning corpus for any IPA-based TTS architecture
that operates at 24 kHz.

---

## Adding a new speaker

The corpus is designed to be extended without breaking existing data. To add
a new speaker (e.g. `spk_m3`, `id=4`):

1. Run the preparation notebook with `TARGET_SPK = 'spk_m3'` and `SPEAKER_ID = 4`.
2. New chunks will be written as `chunk_spk_m1_{index:05d}.parquet` — a different
   prefix from existing speakers, so no overwrite is possible.
3. Training notebooks filter by `speaker_name` before downloading, so they
   only pull the chunks they need.

---

## License

`cc-by-nc-sa-4.0` — non-commercial use, attribution required, share-alike.
Before commercial use, verify the licensing of the source text
(`mimba/text2text`) and OmniVoice-generated audio independently.

---

## Citation

If you use this corpus, please credit the **Mimba** project and cite the
source datasets:

```bibtex
@misc{mimba2026styletts2pltcorpus,
  title   = {Mimba StyleTTS2 PLT Corpus: A Plateau Malagasy TTS Training Corpus},
  author  = {Mimba Ngouana Fofou},
  year    = {2026},
}
```

### Contact

For questions or contributions, open a discussion in the "Community" tab of
this repository.

##### *Contact: [@Mimba](baounabaouna@gmail.com)*