README.md

---
language:
- ka
license: cc-by-nc-4.0
base_model: SWivid/F5-TTS
tags:
- tts
- text-to-speech
- georgian
- f5-tts
- speech-synthesis
- flow-matching
pipeline_tag: text-to-speech
datasets:
- NMikka/Common-Voice-Geo-Cleaned
---

# F5-TTS Georgian

A fine-tuned version of [SWivid/F5-TTS](https://huggingface.co/SWivid/F5-TTS) (335M params) for **Georgian text-to-speech**. The model produces high-quality Georgian speech when using training speakers as reference. Generalization to arbitrary voice cloning is a work in progress.

## Model Details

| | |
|---|---|
| **Base model** | [SWivid/F5-TTS v1 Base](https://huggingface.co/SWivid/F5-TTS) (335M params, DiT + ConvNeXt V2) |
| **Fine-tuning** | Full fine-tune (continuation of flow-matching pretraining), no LoRA |
| **Training data** | [NMikka/Common-Voice-Geo-Cleaned](https://huggingface.co/datasets/NMikka/Common-Voice-Geo-Cleaned) — 20,300 samples, 12 speakers |
| **Training** | 110,000 updates (~100 epochs), single NVIDIA RTX A6000 (48GB) |
| **Sample rate** | 24 kHz |
| **Voice cloning** | Works well with training speakers; generalizing to new voices is WIP |
| **License** | CC-BY-NC-4.0 (inherited from F5-TTS pretrained weights) |

## Evaluation — FLEURS Georgian Benchmark (979 unseen samples)

Round-trip CER: TTS generates audio → [Meta Omnilingual ASR 7B](https://huggingface.co/facebook/omniASR_LLM_7B) transcribes → compare to original text.

| Metric | Value |
|---|---|
| **CER mean** | **0.0509** |
| CER median | 0.0309 |
| CER p90 | 0.1183 |
| CER std | 0.0558 |
| WER mean | 0.1866 |
| WER median | 0.1600 |

**CER distribution:**
- 65.9% of samples < 5% CER
- 85.9% of samples < 10% CER
- 96.5% of samples < 20% CER
- 0 catastrophic failures (> 50% CER)

Evaluated with speaker 3 reference audio (NISQA MOS 4.99).

## Usage

### Install

```bash
pip install f5-tts
```

### Download Model

```python
from huggingface_hub import hf_hub_download

# Download checkpoint and vocab
ckpt_path = hf_hub_download("NMikka/F5-TTS-Georgian", "model_110000.pt")
vocab_path = hf_hub_download("NMikka/F5-TTS-Georgian", "extended_vocab.txt")
```

### Inference

The model works best with reference audio from the training dataset. Voice cloning to arbitrary Georgian speakers is a work in progress.

```python
from datasets import load_dataset
from huggingface_hub import hf_hub_download
from f5_tts.api import F5TTS
import soundfile as sf
import numpy as np

# Download model
ckpt_path = hf_hub_download("NMikka/F5-TTS-Georgian", "model_110000.pt")
vocab_path = hf_hub_download("NMikka/F5-TTS-Georgian", "extended_vocab.txt")

# Load a reference sample from the training dataset
ds = load_dataset("NMikka/Common-Voice-Geo-Cleaned", split="test")
ref_sample = ds[0]  # Pick any sample as voice reference

# Save reference audio to temp file (F5-TTS expects a file path)
ref_path = "/tmp/ref.wav"
sf.write(ref_path, np.array(ref_sample["audio"]["array"]), ref_sample["audio"]["sampling_rate"])

# Load model
model = F5TTS(
    ckpt_file=ckpt_path,
    vocab_file=vocab_path,
    device="cuda",
    use_ema=False,  # Important: this checkpoint was not trained with EMA
)

# Generate speech using a training speaker as reference
wav, sr, _ = model.infer(
    ref_file=ref_path,
    ref_text=ref_sample["text"],
    gen_text="გამარჯობა, როგორ ხარ? საქართველო ულამაზესი ქვეყანაა.",
)
sf.write("output.wav", wav, sr)
```

### Generation Parameters

```python
wav, sr, _ = model.infer(
    ref_file="reference.wav",
    ref_text="reference transcript",
    gen_text="text to synthesize",
    nfe_step=32,       # Denoising steps (default 32, higher = better quality, slower)
    cfg_strength=2.0,  # Classifier-free guidance (default 2.0)
    speed=1.0,         # Speech speed multiplier
)
```

## Training Details

| | |
|---|---|
| **Method** | Full fine-tune (flow-matching loss, continuation of pretraining) |
| **Base checkpoint** | `F5TTS_v1_Base/model_1250000.safetensors` |
| **Learning rate** | 1e-5 |
| **Warmup** | 500 steps |
| **Batch size** | 9,600 audio frames per GPU |
| **Max sequences/batch** | 64 |
| **Optimizer** | 8-bit Adam (bitsandbytes) |
| **Epochs** | 100 |
| **Total updates** | 110,000 |
| **Tokenizer** | Character-level (`char`, not `pinyin`) |
| **Vocab** | 2,579 tokens (2,545 pretrained + 34 Georgian characters) |
| **GPU** | 1x NVIDIA RTX A6000 (48GB) |

### Vocab Extension

The pretrained F5-TTS uses a pinyin-based vocabulary (2,545 tokens). For Georgian, we extended the vocabulary by appending 34 Georgian Unicode characters (ა-ჰ + „). New embeddings were initialized with the mean of existing pretrained embeddings, then the text embedding layer was resized from 2,546 → 2,580 dimensions.

## Limitations and Future Work

- **License**: CC-BY-NC-4.0 — non-commercial use only (inherited from F5-TTS weights)
- **Voice cloning to new speakers is limited** — the model clones training speakers well but does not yet generalize to arbitrary Georgian voices. This is an active area of improvement.
- Trained on 12 speakers from Common Voice Georgian — limited speaker diversity
- Some complex Georgian text with rare characters may produce higher error rates
- No emotion or prosody control beyond what the reference audio provides

## Part of the Georgian TTS Benchmark

This model was trained as part of the first Georgian TTS benchmark — a comparative study of 6 open-source TTS architectures. See the full project: [github.com/NMikaa/TTS_pipelines](https://github.com/NMikaa/TTS_pipelines)

## Citation

```bibtex
@misc{f5tts-georgian-2026,
  title={F5-TTS Georgian: Fine-tuned Flow-Matching TTS for Georgian},
  author={NMikka},
  year={2026},
  url={https://huggingface.co/NMikka/F5-TTS-Georgian}
}
```