Bambara VITS

RobotsMali/bam-vits is an experimental multi-speaker Bambara (Bamanankan, bm) text-to-speech checkpoint from RobotsMali AI4D Lab. It adapts VITS from ylacombe/vits-vctk-with-discriminator, an English VCTK checkpoint, to Bambara.

Research checkpoint — substantially undertrained. This was produced as RobotsMali's first TTS experiment. It was trained for 200 epochs on a small, noisy subset. VITS systems are normally trained for hundreds of thousands of optimizer steps; RobotsMali uses roughly 200,000 steps as a useful reference budget. This run is far below that scale. Expect unstable pronunciation, noise, unnatural prosody, speaker leakage, and occasional unintelligible output. No objective or human evaluation is available.

Model details

Architecture: VITS with stochastic duration prediction and a HiFi-GAN-style waveform decoder
Checkpoint: inference-only transformers.VitsModel; the discriminator was removed
Input: lowercased Bambara in Latin orthography; no phonemizer or uroman step
Audio: 22,050 Hz
Speakers: 20 learned embeddings
Code: RobotsMali-AI/vits-bam, adapted from ylacombe/finetune-hf-vits
Paper: Kim, Kong, and Son (2021)

The project upgrades the upstream dependencies while keeping its core training approach largely unchanged.

Usage

pip install "transformers>=5.9" torch soundfile

import soundfile as sf
import torch
from transformers import AutoTokenizer, VitsModel

repo_id = "RobotsMali/bam-vits"
text = "An ka taa sugu la."
speaker_id = 0  # valid range: 0 through 19

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = VitsModel.from_pretrained(repo_id)
inputs = tokenizer(text.lower(), return_tensors="pt")

with torch.no_grad():
    waveform = model(**inputs, speaker_id=speaker_id).waveform[0]

sf.write("bambara.wav", waveform.cpu().numpy(), model.config.sampling_rate)

The training script remapped participant identifiers to contiguous indices without exporting that mapping. Model IDs 0–19 therefore cannot be reliably mapped back to named AfVoices participants. Audition them for research; do not present an ID as a verified identity.

Training

The model used top20-speakers from RobotsMali/afvoices-notag: 21,253 train and 1,128 test examples from the 20 AfVoices participants with the most utterances, excluding transcripts with semantic/acoustic tags. AfVoices is spontaneous, variably noisy speech collected for ASR rather than studio TTS.

The reproducible configuration is config/bam-vits.yaml.

Setting	Value
Epochs	200
Per-device train batch size	80
Learning rate	0.0005
Precision	FP16
Audio duration filter	0.2–20 s
Maximum token length	450
Seed	789

Vocabulary and speaker embeddings were resized. Text was lowercased and audio resampled to 22.05 kHz.

Evaluation and experimental finding

No MOS, intelligibility, pronunciation, speaker-similarity, or automated TTS results are reported. The held-out split was used during training, but losses are not perceptual evaluation.

Compared informally with bam-vits-pseudo-ipa, pseudo-IPA produced no remarkable improvement in quality or convergence; plain Bambara often sounded slightly more natural. RobotsMali's working explanation is that Bambara orthography is already largely phonetic and that the large English–Bambara acoustic mismatch gave the transferred waveform generator little useful English-phonetic guidance. This is an observation from an undertrained experiment, not a controlled conclusion.

Intended use, risks, and limitations

This release is for low-resource TTS research, listening experiments, baselines, and further fine-tuning. It is not production quality. The limited, non-studio corpus can cause noise, disfluencies, pronunciation errors, demographic imbalance, and speaker leakage. Numbers, abbreviations, foreign words, code-switching, unusual punctuation, long text, and non-Bambara input may fail. Safety, bias, memorization, and voice similarity have not been systematically evaluated.

Do not use it where intelligibility or identity is safety-critical, to impersonate a person, or to create deceptive audio. Use short Bambara sentences, listen critically, and disclose that output is synthetic and experimental.

Related checkpoints

bam-vits-train: discriminator retained for continued training
bam-vits-fintech: adapted on single-speaker FinBamSpeech
bam-vits-pseudo-ipa: pseudo-IPA experiment

Citation

Please cite VITS and identify this checkpoint by repository ID.

@inproceedings{kim2021vits,
  title={Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech},
  author={Kim, Jaehyeon and Kong, Jungil and Son, Juhee},
  booktitle={Proceedings of the 38th International Conference on Machine Learning},
  pages={5530--5540},
  year={2021}
}

Questions are welcome in the project repository or this model's Community tab.