Bambara VITS — pseudo-IPA experiment

RobotsMali/bam-vits-pseudo-ipa is an experimental 20-speaker Bambara VITS checkpoint from RobotsMali AI4D Lab. It differs from RobotsMali/bam-vits in its text input representation, not its VITS architecture: Bambara spelling is converted by a deterministic pseudo-IPA cleaner before tokenization.

Research checkpoint — substantially undertrained. The model is substantially undertrained. It received 200 epochs on a small, noisy corpus, far below the hundreds of thousands of optimizer steps normally used for VITS (approximately 200,000 is our reference budget). No formal evaluation is available; expect unstable, noisy, or unintelligible output.

Research question and finding

The experiment tested whether IPA-like inputs would improve quality or convergence when adapting an English checkpoint trained with English phonetic inputs. We observed no remarkable improvement. Plain-orthography checkpoints often sounded slightly more natural. Bambara orthography is already largely phonetic, and the large Bambara–English acoustic difference likely gave the transferred HiFi-GAN-style generator little useful guidance from either spelling scheme. This is an informal result from undertrained models, not a controlled conclusion.

Usage

The tokenizer does not perform pseudo-IPA conversion itself. Apply the same function used in training before every inference request:

import re


def clean_bambara_pseudo_ipa(text):
    """
    Translates pure Bambara text into pseudo-IPA form.
    Normalizes nasalizations and specific consonants (c -> tʃ, j -> dʒ).
    """
    if not text or not isinstance(text, str):
        return ""

    text = text.lower()

    nasal_map = {
        "aan": "ãã", "ɛɛn": "ɛ̃ɛ̃", "een": "ẽẽ", "iin": "ĩĩ",
        "ɔɔn": "ɔ̃ɔ̃", "oon": "õõ", "uun": "ũũ", "an": "ã",
        "ɛn": "ɛ̃", "en": "ẽ", "in": "ĩ", "ɔn": "ɔ̃", "on": "õ",
        "un": "ũ",
    }
    for ortho, ipa in nasal_map.items():
        text = re.sub(rf"{ortho}(?![aeɛioɔu])", ipa, text)

    sound_alignments = {"j": "dʒ", "c": "tʃ", "ɲ": "ɲ", "ŋ": "ŋ"}
    for b_char, ipa_char in sound_alignments.items():
        text = text.replace(b_char, ipa_char)

    text = re.sub(r"[^a-zɛɔɲŋãẽĩõṹ̀̂̌̄\s.,!?ʃʒ]", "", text)
    return re.sub(r"\s+", " ", text).strip()

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

repo_id = "RobotsMali/bam-vits-pseudo-ipa"
text = clean_bambara_pseudo_ipa("An ka taa sugu la.")
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = VitsModel.from_pretrained(repo_id)
inputs = tokenizer(text, return_tensors="pt")

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

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

Use speaker IDs 0–19. Their mapping to original AfVoices participant IDs was not exported and should not be treated as identity metadata.

Training

The model was initialized from ylacombe/vits-vctk-with-discriminator and trained on the 21,253-train/1,128-test top20-speakers split of RobotsMali/afvoices-notag. This is spontaneous, variably noisy ASR speech, not studio TTS data. Transcripts with semantic/acoustic tags were excluded.

The configuration is config/bam-vits-pseudo-ipa.yaml: 200 epochs, per-device batch size 80, learning rate 0.0005, FP16, 0.2–20-second audio, maximum 450 tokens, and seed 789. Audio was resampled to 22.05 kHz. The published checkpoint is inference-only VitsModel; its discriminator was removed.

Intended use and limitations

This is a baseline for research into Bambara text representations, low-resource transfer, and continued fine-tuning—not a production voice. No MOS, intelligibility, pronunciation, speaker-similarity, safety, bias, memorization, or voice-similarity evaluation was performed. The dataset and short training can yield noise, poor prosody, pronunciation errors, speaker leakage, and demographic imbalance. The cleaner is only a heuristic: it is not a linguistic IPA transcription or a complete grapheme-to-phoneme system. Numbers, abbreviations, code-switching, foreign words, unusual punctuation, and long/non-Bambara text may fail.

Do not use the model for safety-critical speech, impersonation, or deceptive audio. Disclose synthetic use and verify all output by listening.