AIHubSN/Kiriku-Pulaar-TTS

Model Card: PulaarTTS

Model Details

Name	pulaarTTS
Architecture	VITS (Conditional Variational Autoencoder with adversarial learning)
Framework	Coqui TTS
Task	Text-to-Speech (TTS) synthesis
Language	pulaar (ff)
Sample rate	22,050 Hz
Checkpoint size	~952 MB (model.pth)
Files	model.pth (weights), config.json (architecture + training config)
Vocabulary	93 grapheme characters (Latin alphabet + pulaar/French diacritics: à, é, ë, ñ, etc.) — no phonemizer, trained directly on graphemes

The model is a single-speaker acoustic model + vocoder trained end-to-end (VITS jointly learns the text-to-spectrogram and spectrogram-to-waveform steps, so no separate vocoder is needed at inference time).

Intended Use

Synthesize spoken pulaar audio from raw pulaar text for the Speech Studio backend.

Not intended for: languages other than pulaar, multi-speaker synthesis, or real-time/streaming use cases (inference generates a full .wav file per request).

How to Use

Install dependency

pip install transformers==4.46.3 TTS==0.19.0

Download Weights From HF

from huggingface_hub import snapshot_download
snapshot_download(repo_id="AIHubSN/Kiriku-Pulaar-TTS", local_dir="checkpoints")

Generate Speech with Coqui TTS

from TTS.utils.synthesizer import Synthesizer

synthesizer = Synthesizer(
    tts_checkpoint="checkpoints/model.pth",
    tts_config_path="checkpoints/config.json",
)

wav = synthesizer.tts(text="Aliw ina jiira kodde, moɗi tan, ɗeɗɗii.")
synthesizer.save_wav(wav, "output.wav")

Training Data

• Dataset: pulaar_dataset/ with a Coqui-formatted metadata.txt transcript file (path/text pairs).
• Audio preprocessing: silence trimmed at 45 dB, 22,050 Hz sample rate, 80 mel bins, no pitch/energy conditioning.
• No phonemizer used — trained on raw graphemes with cleaner disabled (text_cleaner: null), so text is passed through largely as-is (after lower-casing/number conversion at inference time).
• Dataset size, number of speakers/hours, and licensing are not recorded in config.json — fill this in if known before distributing the card externally.

Training Procedure

• Base config: single-speaker VITS (use_speaker_embedding: false, num_speakers: 0)
• Epochs: up to 2000 (checkpoints saved every 12,000 steps, best-3 kept)
• Batch size: 32
• Optimizer: AdamW (β=[0.8, 0.99], eps=1e-9, weight_decay=0.01)
• Learning rates: generator/discriminator both start at 2e-4, ExponentialLR decay (γ=0.999875/epoch)
• Loss weights: mel=45.0, KL=1.0, discriminator/generator/feature/duration=1.0
• Restored from checkpoint TTS_MODELING/pulaarTTS-May-31-2026_04+35PM-.../best_model_2017892.pth (~2M training steps)

Evaluation

No quantitative metrics (MOS, MCD, WER, etc.) are recorded alongside the checkpoint. Qualitative test sentences used during training (for manual listening checks):

• "ko ɗuum waɗi ɓe kuccini bonanndeeji maɓɓe ko he sektoraaji jogiiɗi nguurndam leydi ndi ɗi"
• "Hayso tawii ko Makki Sal heddoraa he hoore laamu leydi Senegaal natti heewde ne"
• "ko wayno isinaaji ndiyam e jeyngol kam e jeereeji"

If you need a documented quality bar, run these (or a held-out eval set) through the model and record MOS/CMOS results here.

Limitations & Risks

• Low-resource language: Pulaar TTS training data is comparatively small/scarce versus high-resource languages; expect more artifacts, mispronunciations on out-of-distribution text, and weaker prosody on long or rare sentences.
• Single speaker, single style: no control over speaker identity, emotion, or speaking rate beyond VITS's built-in length_scale/noise parameters.
• Grapheme-based, no phonemizer: pronunciation of unseen character sequences (loanwords, typos, French-mixed text beyond digits) may be unreliable.
• No content filtering: the model will synthesize whatever text it's given; add moderation upstream if used in a user-facing product.