Bambara VITS — training checkpoint
RobotsMali/bam-vits-train is the continued-training counterpart of RobotsMali/bam-vits. Its VitsModelForPreTraining weights retain both the generator and discriminator (about 323 MB rather than the inference export's 159 MB). Use this repository as model_name_or_path when fine-tuning; use bam-vits for ordinary inference.
Substantially undertrained. The checkpoint received 200 epochs on a small experimental corpus and remains far below typical VITS budgets of hundreds of thousands of optimizer steps (approximately 200,000 is our reference). It is a starting point, not a converged base model.
Continue training
Clone the RobotsMali training code, install its pinned dependencies, build monotonic alignment, then point a copied YAML config at this checkpoint and your dataset:
git clone https://github.com/RobotsMali-AI/vits-bam.git
cd vits-bam
python -m pip install -r requirements.txt
cd monotonic_align && python setup.py build_ext --inplace && cd ..
cp config/bam-vits.yaml config/my-experiment.yaml
At minimum, edit:
model:
model_name_or_path: RobotsMali/bam-vits-train
tokenizer_name: RobotsMali/bam-vits-train
pseudo_ipa: false
override_speaker_embeddings: true
override_vocabulary_embeddings: false
data:
dataset_name: OWNER/DATASET
dataset_config_name: null
audio_column_name: audio
text_column_name: text
speaker_id_column_name: speaker_id
logging:
output_dir: ./tmp/my-experiment
hub_model_id: OWNER/MODEL
Then run:
accelerate launch run_vits_finetuning.py config/my-experiment.yaml
Review speaker and vocabulary settings for your data. This repository's custom training class and compiled monotonic-alignment extension are required. After training, push_vits_model.py --inference_only publishes a generator-only VitsModel; omitting the flag preserves the discriminator.
Provenance
The model was initialized from the English VCTK ylacombe/vits-vctk-with-discriminator and trained on top20-speakers from RobotsMali/afvoices-notag: 21,253 training and 1,128 test examples from the 20 participants with the most utterances, excluding tagged transcripts. The corpus is spontaneous, noisy ASR data rather than studio TTS data.
The source config, config/bam-vits.yaml, specifies 200 epochs, per-device batch 80, learning rate 0.0005, FP16, 0.2–20-second audio, maximum 450 tokens, and seed 789. Text was lowercased; audio was resampled to 22.05 kHz.
Limitations and responsible use
No MOS, intelligibility, pronunciation, speaker-similarity, or safety evaluation is reported. Any downstream model inherits risks from undertraining, noisy data, limited speaker coverage, minimal text normalization, and the English initialization. The source participant-to-model-index mapping was not saved. Evaluate convergence, pronunciation, memorization, bias, licensing, speaker consent, and voice similarity for every derivative.
This checkpoint is for research and fine-tuning. Do not use it to impersonate people, create deceptive audio, or support safety-critical speech without substantial additional training and evaluation. Clearly document downstream data and disclose synthetic audio.