MayaD8/omnivoice-georgian

OmniVoice Georgian

A Georgian fine-tune of k2-fsa/OmniVoice (Qwen3-0.6B + Higgs Audio V2 codec), a zero-shot voice-cloning text-to-speech model.

Training

• Base model: k2-fsa/OmniVoice
• Data: ~88.8 hours from a private read-speech corpus (Chapidze), duration-consistency filtered (dropped ~1.4% of clips as chars/sec IQR outliers — likely misaligned transcripts), on top of a starting checkpoint that also includes Common Voice Georgian
• Steps: 1,900 (~2 epochs), continued fine-tuning from our own prior Georgian checkpoint
• Learning rate: 2e-5, cosine schedule
• Final eval loss: ~4.52 (best checkpoint during the run reached ~4.49)
• Hardware: single NVIDIA L4 (24GB)
• Precision: bf16, sdpa attention

This is one of several experiments trying to fix audible pausing/pacing glitches in an earlier version trained from scratch on a noisier 278.8h mix (which included broadcast/meeting-room recordings). This version instead builds on our own already-adapted Georgian checkpoint and adds only Chapidze, to isolate whether Chapidze itself is the source of the artifact.

Usage

pip install git+https://github.com/k2-fsa/OmniVoice.git

from omnivoice import OmniVoice
import soundfile as sf

model = OmniVoice.from_pretrained("MayaD8/omnivoice-georgian").to("cuda")

# Auto mode - model picks a voice
audios = model.generate(text="გამარჯობა, როგორ ხარ?", language="ka")
sf.write("out.wav", audios[0], model.sampling_rate)

# Voice cloning mode - clone from a reference clip
audios = model.generate(
    text="გამარჯობა, როგორ ხარ?",
    language="ka",
    ref_audio="reference.wav",
    ref_text="reference audio's transcript",
)
sf.write("out_cloned.wav", audios[0], model.sampling_rate)

License

Apache 2.0, inherited from the base k2-fsa/OmniVoice model.