rnjema-unima/mms-tts-fat-baseline

WAXAL MMS-TTS — Akan – Fante (fat)

Fine-tuning-ready checkpoint for Akan – Fante (fat).

WAXAL dataset config	google/WaxalNLP — fat_tts
Data provider	University of Ghana
WAXAL data license	CC-BY-4.0
Base model	facebook/mms-tts-aka
Model license	CC-BY-NC 4.0 (MMS base; governs fine-tuned model)

⚠️ Proxy checkpoint

facebook/mms-tts-fat does not exist in MMS-TTS coverage (1107 languages). This repository fine-tunes from the closest available linguistic donor:

Proxy donor	facebook/mms-tts-aka
Ranked alternatives	aka
Other WAXAL languages sharing this donor	twi (Akan – Twi), bci (Baoulé)

Proximity rationale: Akan-Fante (fat) has no MMS checkpoint. Proxy: aka (Akan macro-language, MMS). Fante is a dialect of Akan; aka covers the full Akan continuum including Fante. fat, twi, and bci all use aka as their proxy donor.

Each recipient language is fine-tuned independently from the same donor base. Donor weights provide acoustic/prosodic warm-start; WAXAL fine-tuning adapts them to the target language.

What this repository adds

facebook/mms-tts-* Hub checkpoints are inference-only releases that crash run_vits_finetuning.py. This repository applies three patches:

File	Change
config.json	pad_token_id set to 0 (was null)
tokenizer_config.json	pad_token entry added
preprocessor_config.json	Added — VitsFeatureExtractor config from ylacombe/mms-tts-eng-train

Model weights are not stored here. _name_or_path in config.json points to facebook/mms-tts-aka, so run_vits_finetuning.py loads weights from that checkpoint at training time.

preprocessor_config.json

Downloaded verbatim from ylacombe/mms-tts-eng-train. Values are VITS architecture constants shared by all MMS-TTS languages.

Field	Value
feature_extractor_type	VitsFeatureExtractor
feature_size	80
hop_length	256
max_wav_value	32768.0
n_fft	1024
padding_side	right
padding_value	0.0
return_attention_mask	False
sampling_rate	16000
spec_gain	1

Usage in finetune-hf-vits

{
  "model_name_or_path":     "rnjema-unima/mms-tts-fat-baseline",
  "feature_extractor_name": "rnjema-unima/mms-tts-fat-baseline",
  "dataset_name":           "google/WaxalNLP",
  "dataset_config_name":    "fat_tts",
  "audio_column_name":      "audio",
  "text_column_name":       "text",
  "train_split_name":       "train",
  "eval_split_name":        "validation"
}

Inference (after fine-tuning)

from transformers import VitsModel, VitsTokenizer
import torch, scipy

model     = VitsModel.from_pretrained("your-org/your-finetuned-model")
tokenizer = VitsTokenizer.from_pretrained("your-org/your-finetuned-model")

inputs = tokenizer("Your text in Akan – Fante.", return_tensors="pt")
with torch.no_grad():
    out = model(**inputs)

scipy.io.wavfile.write("output.wav", model.config.sampling_rate,
    out.waveform.squeeze().numpy())

Technical details

Architecture	VITS (end-to-end, no separate vocoder)
MMS match type	proxy
pad_token_id	0
vocab_size	30
is_uroman	false
sampling_rate	16000 Hz

References

BibTex Citation:

This model was developed by Vineel Pratap et al. from Meta AI. If you use the model, consider citing the MMS paper:

@article{pratap2023mms, title={Scaling Speech Technology to 1,000+ Languages}, author={Vineel Pratap and Andros Tjandra and Bowen Shi and Paden Tomasello and Arun Babu and Sayani Kundu and Ali Elkahky and Zhaoheng Ni and Apoorv Vyas and Maryam Fazel-Zarandi and Alexei Baevski and Yossi Adi and Xiaohui Zhang and Wei-Ning Hsu and Alexis Conneau and Michael Auli}, journal={arXiv}, year={2023} }

Notable Acknowledgements:

• WAXAL paper — arXiv:2602.02734
• WAXAL dataset
• finetune-hf-vits
• facebook/mms-tts-aka
• preprocessor_config.json source — ylacombe/mms-tts-eng-train