Datasets:
Mimba OmniVoice Voice Bank — Reusable Multi-Speaker Reference Voices
A bank of distinct synthetic reference voices, generated with OmniVoice Voice Design (attribute-driven generation, no reference audio). Each entry is a short reference clip + its attributes + a speaker embedding.
The bank is language-agnostic: it captures only the timbre of each voice. It
is meant to be reused as ref_audio to generate multi-speaker TTS corpora in any
language (Plateau Malagasy plt, Ngiemboon nnh, …), so that a cloning model
learns to separate content from speaker across hundreds of timbres instead of a
handful — the key factor for robust zero-shot voice cloning.
▶️ Audio plays directly in the dataset viewer (the
ref_audiocolumn is a standard 🤗Audiofeature). Click any row to listen.
⚠️ Synthetic, not human. All audio is generated by OmniVoice. The reference sentence is in English (used only to capture timbre). See Limitations.
Dataset summary
| Property | Value |
|---|---|
| Generator | OmniVoice Voice Design (instruct=, no reference audio) |
| Voices generated | 1000 (all attribute combinations) |
| Unique voices after dedup | 527 (ECAPA/Resemblyzer cosine > 0.95 removed) |
| Gender split (unique) | 304 male · 223 female |
| Audio | Mono, 24 000 Hz, PCM-16 WAV bytes embedded in Parquet |
| Reference text | Single English sentence (timbre capture only) |
| Selection list | voices_clean.csv — the deduplicated set of unique voices |
The shards (data/voicebank-*.parquet) contain all 1000 generated voices
(so you can re-run dedup with your own threshold). The curated 527 unique voices
are listed in voices_clean.csv (metadata only; join on name to get audio).
Attribute grid (Voice Design)
{gender} × {age} × {pitch} × {style} × {accent} → naming {gender}_{age}_{pitch}_{style}_{accent}.
| Category | Values |
|---|---|
| Gender | male, female |
| Age | child, teenager, young adult, middle-aged, elderly |
| Pitch | very low, low, moderate, high, very high |
| Style | normal, whisper |
| Accent | american, british, australian, canadian, indian, chinese, korean, japanese, portuguese, russian |
⚡ How to use
Listen / browse
from datasets import load_dataset
ds = load_dataset("mimba/omnivoice-voicebank", split="train")
row = ds[0]
print(row["name"], "|", row["gender"], row["age"], row["pitch"], row["accent"])
audio = row["ref_audio"] # {'array': np.float32[...], 'sampling_rate': 24000}
# jouer dans un notebook
import IPython.display as ipd
ipd.display(ipd.Audio(audio["array"], rate=audio["sampling_rate"]))
Use the deduplicated selection (recommended)
import pandas as pd
from huggingface_hub import hf_hub_download
clean = pd.read_csv(hf_hub_download("mimba/omnivoice-voicebank", "voices_clean.csv",
repo_type="dataset"))
# ex. exclure les chuchotements (indistinguables entre eux) pour l'entraînement :
clean = clean[clean["style"] != "whisper"]
print(len(clean), "voix uniques utilisables")
Reuse a voice as a cloning reference (any language)
import io, soundfile as sf
from datasets import load_dataset
ds = load_dataset("mimba/omnivoice-voicebank", split="train")
row = next(r for r in ds if r["name"] == "female_youngadult_high_normal_british")
sf.write("ref.wav", row["ref_audio"]["array"], row["ref_audio"]["sampling_rate"])
# -> passe ref.wav comme ref_audio à ton TTS/VC, avec le texte de la langue cible.
Data fields
| Field | Type | Description |
|---|---|---|
name |
string | {gender}_{age}_{pitch}_{style}_{accent} |
gender, age, pitch, style, accent |
string | Attributs Voice Design |
instruct |
string | Chaîne instruct= exacte envoyée à OmniVoice |
ref_text |
string | Phrase anglaise de référence (capture du timbre) |
sample_rate |
int64 | 24000 |
ref_audio |
audio | {bytes: <WAV 24 kHz mono PCM-16>, path: None} — joué par le viewer |
embedding |
list[float32] | Embedding locuteur (Resemblyzer) — sert à la dédup |
How it was produced
- Generation — 1000 voix via OmniVoice Voice Design (2×5×5×2×10 combinaisons
d'attributs), texte de référence anglais unique,
num_step=32, déterministe. - Filtrage qualité — rejet des audios silencieux (RMS < seuil) ou trop courts.
- Déduplication — embedding locuteur (Resemblyzer) par voix ; suppression
gloutonne des voix dont la similarité cosinus dépasse 0.95 (une gardée par
grappe). Résultat : 527 voix uniques listées dans
voices_clean.csv. - Stockage — shards Parquet (
data/voicebank-XXXX.parquet), audio en bytes WAV 24 kHz, push incrémental + purge disque locale.
Limitations
- Synthétique, pas humain. Le timbre et la prosodie héritent d'OmniVoice.
- Chuchotements peu distinguables. Les voix
whisperse ressemblent fortement pour les encodeurs de locuteur (parole non voisée) → elles sont massivement fusionnées à la dédup ; les quelqueswhisperconservées sont distinctes, mais pour l'entraînement il est recommandé de filtrerstyle == "whisper". - Audio de référence en anglais. Seul le timbre est réutilisable ; la langue du contenu cible se choisit au moment de la génération (clonage).
- Labels d'attributs = intention, pas garantie. Certaines combinaisons
(ex.
child+very low pitch) sont approximées par le modèle.
Intended use
Enrichir la diversité de locuteurs lors du finetuning de TTS/clonage sur langues
peu dotées (projet Mimba), afin de préserver la généralisation zero-shot du
clonage quand les données natives ne comptent que quelques locuteurs. Réutilisable
comme banque de références pour plt, nnh, etc.
License
cc-by-nc-sa-4.0 — usage non commercial, attribution, partage à l'identique.
Vérifier indépendamment la licence d'OmniVoice avant tout usage commercial.
Citation
@misc{mimba2026omnivoicevoicebank,
title = {Mimba OmniVoice Voice Bank: Reusable Multi-Speaker Reference Voices},
author = {Mimba Ngouana Fofou},
year = {2026},
}
Contact: @Mimba
- Downloads last month
- 34