F5-TTS Hungarian — Native Magyar Text-to-Speech with Zero-Shot Voice Cloning

The first and only Hungarian fine-tuned F5-TTS model with zero-shot voice cloning from just 5-15 seconds of reference audio.

  Training on H100 GPUs is expensive — a coffee directly funds the next epochs. 🙏

Highlights

• Native Hungarian — Trained on ~145,000 Hungarian speech samples across three datasets (~270 hours)
• Zero-shot Voice Cloning — Clone any voice with just 5-15 seconds of reference audio
• Runs Locally — No cloud API needed, runs on consumer GPUs (8GB+ VRAM)
• Fast Inference — Real-time factor < 0.5 on RTX 3060+
• Privacy First — All processing happens on your machine

Audio Samples

Listen to what the model can do based on 5-15 seconds of Hungarian reference audio!

Model Details

Training Datasets

Training Parameters

Quick Start

import torch
import torchaudio
import soundfile as sf
import numpy as np

# Monkey-patch torchaudio for cross-platform compatibility
_orig_load = torchaudio.load
def _patched_load(fp, **kw):
    d, sr = sf.read(str(fp), dtype="float32")
    if d.ndim == 1: d = d[np.newaxis, :]
    else: d = d.T
    return torch.from_numpy(d), sr
torchaudio.load = _patched_load

from f5_tts.api import F5TTS

model = F5TTS(
    model="F5TTS_v1_Base",
    ckpt_file="model_last_final.safetensors",
    vocab_file="vocab.txt",
    device="cuda",
    use_ema=True,
)

wav, sr, _ = model.infer(
    ref_file="your_reference.wav",
    ref_text="A referencia hang pontos atirata.",
    gen_text="Szia, ez egy teszt mondat a magyar szovegfelolvasashoz.",
)

sf.write("output.wav", wav, sr)

Critical: Reference Text Must Match Audio

The ref_text must be the exact transcription of ref_audio. Mismatched text causes garbled output. Use Whisper to transcribe:

from faster_whisper import WhisperModel
whisper = WhisperModel("large-v3-turbo", device="cuda")
segments, _ = whisper.transcribe("your_reference.wav", language="hu")
ref_text = " ".join(s.text.strip() for s in segments)

Benchmark

Measured on NVIDIA RTX 5060 Ti 16GB, PyTorch 2.x, CUDA, FP16:

Reference Audio Guidelines

For optimal voice cloning quality:

• Use 5-15 seconds of clean speech (no background noise, no reverb, no music)
• Provide an exact transcript — this is the single most important quality factor
• Mono WAV, any sample rate (automatically resampled to 24kHz)
• Longer references (>15s) do NOT improve quality and increase VRAM usage
• End the reference audio with a natural sentence ending, not mid-word

Repository Contents

Known Issues & Workarounds

Onset Artifact (First-Word Distortion)

The model occasionally produces a brief garbled sound (~200-400ms) at the very beginning of generated audio. This is a known characteristic of the F5-TTS DiT architecture — the hard mel-spectrogram boundary between reference and generated speech does not always align with the acoustic boundary.

Workarounds (in order of effectiveness):

1. Prefix trick — Prepend a short filler word to gen_text (e.g., "szoval, " + your_text), then trim the first ~400ms from the output
2. Silence padding — Add 500-700ms of silence to the end of your reference audio before inference
3. VAD-based trimming — Use Silero VAD to detect actual speech onset and trim everything before it
4. Regeneration — Different random seeds produce different alignment; retry 2-3 times if needed

The included inference_example.py implements an energy-based adaptive trimmer. Community contributions for better solutions are welcome!

Duration Estimation

Very short texts (<5 characters) may produce stretched or compressed audio. The model estimates duration from the ref_text/gen_text character ratio — extremely short inputs can break this heuristic. Workaround: pad short texts with natural filler.

Roadmap

• Improved onset artifact elimination (Silero VAD + soft mel crossfade)
• Streaming inference support (sentence-level chunking)
• ComfyUI custom node
• Companion model: Whisper Hungarian LoRA for STT
• Extended vocabulary with English code-switching support
• INT8 quantized variant for lower VRAM (<4GB)

Community & Contributions

The model weights are released under CC-BY-NC-4.0 (following the base F5-TTS license). The inference code, preprocessing scripts, and tooling are open for contributions!

We welcome:

• Better Hungarian text normalization rules
• Improved onset artifact handling
• New reference audio / evaluation pairs
• Bug reports and quality feedback
• Translations of documentation

GitHub (Inference toolkit & Issues): github.com/Maxdorger/f5-tts-hungarian Hugging Face (Model Weights): huggingface.co/Maxdorger29/f5-tts-hungarian

Support the Project

Training high-quality TTS models on H100 GPUs is expensive. If this model is useful to you, consider buying me a coffee — it directly funds compute time for future improvements!

License

This model is released under CC-BY-NC-4.0, consistent with the base F5-TTS model license.

• Personal & Research use: Free
• Commercial use: Not permitted under this license (upstream restriction from F5-TTS base model)
• Attribution: Credit "F5-TTS Hungarian by Maxdorger29" in any published work

Acknowledgments

• SWivid/F5-TTS — Base model architecture and pretrained weights
• Mozilla Common Voice — Hungarian speech data
• YodaLingua — Hungarian speech dataset
• CSS10 — Hungarian single-speaker data

Citation

@misc{f5tts-hungarian-2026,
  title={F5-TTS Hungarian: Native Magyar Text-to-Speech with Zero-Shot Voice Cloning},
  author={Maxdorger29},
  year={2026},
  url={https://huggingface.co/Maxdorger29/f5-tts-hungarian},
  note={Fine-tuned on ~145k Hungarian speech samples using NVIDIA H100}
}