MODEL CARD — Pokoot West TTS
Base Model
- HuggingFace ID:
facebook/mms-tts-kpz(facebook/mms-tts-kpz) - Architecture: VITS — Variational Inference with adversarial learning for end-to-end TTS
Dataset
- HuggingFace ID:
speedykom-group/pokoot-west-speech-dataset - Splits used: train / validation / test
- Size: see dataset card on HuggingFace Hub for exact utterance counts and total hours
- License of training data: see dataset card on HuggingFace Hub — verify before redistribution
Fine-tuning Details
- Trainable layers: text_encoder
- Optimizer: AdamW (lr=1e-05, betas=(0.8, 0.99), weight_decay=0.01)
- Scheduler: CosineAnnealingLR
- Batch size: 4 samples × 4 accumulation steps = effective batch 16
- Max epochs: 50 (early stopping on validation mel loss, patience=5)
- Best checkpoint: selected by minimum validation mel-spectrogram loss (L1)
Evaluation (Test Set)
Mel-spectrogram loss (L1) and MCD (Mel Cepstral Distortion, dB — lower is better)
Deliverable Files
| File | Description |
|---|---|
config.json |
Model architecture config |
model.safetensors |
Fine-tuned weights (preferred over .bin) |
tokenizer_config.json |
Tokenizer settings |
vocab.json |
Vocabulary mapping |
special_tokens_map.json |
Special token definitions |
samples/ |
5 synthesised WAVs listened to and approved |
Sample Audio
Five synthesised WAVs are included in the samples/ folder.
Each has been listened to and approved before delivery.
| Pokoot West | English | Audio |
|---|---|---|
| Kebaak kotik chebo rap. | Heavy rain is coming. | sample_01.wav |
| Munyek kechop kotik. | Strong winds are expected. | sample_02.wav |
| Kany tumwek chebo ng’wony. | Be careful of flooding. | sample_03.wav |
| Mii ng’alechu ak wuut. | Stay safe at home. | sample_04.wav |
| Tinyei kokwaab chebo barak. | There may be thunderstorms. | sample_05.wav |
Intended Use
Speech synthesis for Pokoot West weather and emergency-alert messages. Not intended for speaker identification or any processing of personal data.
Limitations
- Fine-tuned only on the Pokoot West speech dataset; out-of-vocabulary or code-switched text may produce degraded output.
- Model inherits any biases present in the base MMS-TTS checkpoint and training data.
License
cc-by-nc-4.0 license.
Training data: see dataset card for speedykom-group/pokoot-west-speech-dataset.
- Downloads last month
- 6
Model tree for speedykom-group/mms-tts-pko-west
Base model
facebook/mms-tts-kpz