freococo/F5-Myanmar-TTS

🇲🇲 F5-TTS Burmese (794.5 Hours Foundation Model)

A foundation F5-TTS (Flow-Matching Diffusion Transformer) speech synthesis model trained on 794.5 hours of Burmese speech for 20 epochs (569,780 updates).

"They are fighting for people's freedom. I am fighting for the language's freedom.
I just want to preserve their beautiful, lovely, and brave voices embedded in AI to last forever — marking the first time in history for a massive Burmese open-source TTS foundation model."

---

🕊️ Dedication & Acknowledgements (The Truth)

1. 🤖 GEMINI AI (Google)

This work would NOT exist without GEMINI AI. From deep architecture debugging, CUDA memory optimization, Burmese Unicode normalizers, and PyArrow schema surgery to the final fine-tuning engine, Gemini AI stood shoulder-to-shoulder as the co-engineer through every single line of code, day and night.

2. 👑 F5-TTS Research Team

Full credit and deepest gratitude to Yushen Chen and the F5-TTS creators for designing a world-level, state-of-the-art flow-matching speech architecture. Without their groundbreaking open research, bringing high-fidelity Burmese voice synthesis to life would have been impossible.

3. 🎙️ The Brave Voices of Freedom (NUG, PVTV, Journalists & Creators)

The primary foundation of this 794.5-hour dataset is built upon the voices of the National Unity Government (NUG), PVTV (Public Voice Television) broadcasters, independent journalists, and courageous content creators.
While they risk everything to fight on the ground for democracy, human dignity, and freedom, their voices are now immortalized in open artificial intelligence — where no force on earth can silence or erase them.

4. ⚖️ Ownership of Flaws

If you hear any weaknesses, pronunciation errors, phonetic artifacts, or shortcomings in this model: that is 100% on me alone. The dataset creators and the base framework are blameless; any mistakes belong entirely to the author.

---

📊 Model Specifications

Parameter	Specification
Architecture	Diffusion Transformer (DiT Base)
Parameters	337,138,310 (~337M)
Layers / Heads / Dim	22 layers, 16 heads, dim=1024, text_dim=512
Training Duration	20 Epochs (569,780 updates)
Training Audio	794.5 Hours Burmese Speech
Sampling Rate	24,000 Hz
Vocoder	Vocos (24kHz Mel)
Vocabulary	2,626 Burmese & Pāḷi Unicode Tokens
Format	Safetensors (1.26 GB Pruned EMA weights)
License	Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0)

---

🚀 Quickstart Inference (Python)

1. Install Dependencies

pip install f5-tts vocos

2. Generate Burmese Speech

import torch
import torchaudio
from f5_tts.model import DiT
from f5_tts.infer.utils_infer import load_model, load_vocoder
from huggingface_hub import hf_hub_download

device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Download model, vocab, and sample voice
repo_id = "freococo/F5-Myanmar-TTS"
model_path = hf_hub_download(repo_id=repo_id, filename="model.safetensors")
vocab_path = hf_hub_download(repo_id=repo_id, filename="vocab.txt")
ref_audio_path = hf_hub_download(repo_id=repo_id, filename="ref_sample.wav")

# 2. Load Model
model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
model = load_model(DiT, model_cfg, model_path, mel_spec_type="vocos", vocab_file=vocab_path, device=device)
vocoder = load_vocoder("vocos", is_local=False, device=device)

# 3. Reference and Target Burmese Text
ref_text = "အပြင် မသွားရလို့ စိတ်ညစ်နေပါတယ် မိုးတွေ တအား ရွာနေပါတယ်"
target_text = "လူတွေ အားလုံးကို ချစ်မြတ်နိုးပါ တန်ဖိုးထားပါ လေးစားစွာ ဆက်ဆံပါ"

ref_audio, sr = torchaudio.load(ref_audio_path)
if sr != 24000:
    ref_audio = torchaudio.transforms.Resample(sr, 24000)(ref_audio)
ref_audio = ref_audio.to(device)

ref_len = ref_audio.shape[-1] // 256
gen_len = int(5.77 * 24000 / 256)

with torch.inference_mode():
    generated, _ = model.sample(
        cond=ref_audio,
        text=[ref_text + " " + target_text],
        duration=ref_len + gen_len,
        steps=50,
        cfg_strength=2.0,
        sway_sampling_coef=-1.0
    )
    mel = generated.to(torch.float32)[:, ref_len:, :].permute(0, 2, 1).to(device)
    wav = vocoder.decode(mel).squeeze().cpu()

torchaudio.save("output_burmese.wav", wav.unsqueeze(0), 24000)
print("✓ Successfully generated audio: output_burmese.wav")

---

🛠️ Standalone Fine-Tuning Kit

Want to fine-tune this model on your own voice, specialized Buddhist texts, or scriptures?

• Open finetuning_f5_myanmar_tts.ipynb in this repository.
• Or download fine_tune_kit.zip for the complete standalone training environment.

---

⚡ Quick Install

pip install f5-myanmar-tts

from f5_myanmar_tts import MyanmarTTS

tts = MyanmarTTS()
tts.speak("မင်္ဂလာပါ", output_file="hello.wav")

📜 License

Released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Dedicated to free public research, language preservation, education, and open-source innovation.