pythainlp/MakHamTTS

MakHam (มะขาม) Thai Text-to-Speech

MakHam (มะขาม) is an ultra-lightweight, high-performance open Thai Text-to-Speech (TTS) model family developed by PyThaiNLP.

MakHam combines Kokoro-82M multi-voice architecture with the Paradee 1.1 single-voice distillation framework, supporting accurate 5-tone Thai phonology and standalone ONNX deployment across servers, CPUs, edge devices, and web browsers.

• Online demo: https://github.wannaphong.com/MakHamTTS/webdemo/
• Model Repository: pythainlp/MakHamTTS
• Dataset Repository: pythainlp/MakHamTTS-dataset
• Source Code & Training Suite: GitHub: wannaphong/MakHamTTS

---

Model Highlights

• Ultra-Compact Size: 7.88M parameters (~8.07 MB INT8 ONNX file) — reduced from 325 MB Kokoro-82M.
• Blazing Fast: > 45x Real-Time Factor (RTF) on a single modern CPU thread without requiring GPU acceleration.
• 5-Tone Thai Phonology: Explicit tone representation (→ mid, ↓ low, ↘ falling, ↑ high, ↗ rising) powered by FastThaiG2P with a 115-token Kokoro vocabulary.
• Self-Contained ONNX Graph: The female acoustic style representation is baked directly into the graph. No external .npy voicepack tensor or dynamic tensor slicing is needed at runtime.
• Natural Voice Timbre: Warm, natural Thai female voice trained on verified conversational, news, and everyday spoken Thai datasets (~50 hours).

---

Benchmarks & Specifications

Property	MakHam Paradee (INT8)	MakHam Paradee (FP32)	Kokoro-82M Thai Base
Parameters	7,877,918 (~7.88M)	7,877,918 (~7.88M)	82,000,000 (82M)
Model Size	8.07 MB	33.17 MB	325.5 MB
Architecture	TextStudent (s+mlp) + VDecoder (v2)	TextStudent (s+mlp) + VDecoder (v2)	StyleTTS2 + iSTFTNet
Sample Rate	24,000 Hz	24,000 Hz	24,000 Hz
Vocab Size	115 tokens (5-tone)	115 tokens (5-tone)	115 tokens (5-tone)
Voicepack Required	No (Embedded)	No (Embedded)	Yes (ref_s [510, 1, 256])
Speed (CPU)	~45x Real-Time	~28x Real-Time	~4.5x Real-Time

---

Quick Start & Installation

Requirements

pip install onnxruntime soundfile huggingface_hub
pip install https://github.com/awslabs/FastThaiG2P/archive/refs/heads/main.zip

Python (Direct ONNX Runtime & FastThaiG2P)

import json
import numpy as np
import onnxruntime as ort
import soundfile as sf
from huggingface_hub import hf_hub_download
from fastthaig2p import G2P, ipa_to_kokoro

# 1. Download model and config from Hugging Face
model_path = hf_hub_download(
    repo_id="pythainlp/MakHamTTS",
    filename="paradee_thai/paradee_thai_int8.onnx"
)
config_path = hf_hub_download(
    repo_id="pythainlp/MakHamTTS",
    filename="paradee_thai/config.json"
)

# 2. Initialize G2P and ONNX session
g2p = G2P()
with open(config_path, "r", encoding="utf-8") as f:
    cfg = json.load(f)

session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])

# 3. Convert Thai text to Kokoro phoneme IDs
text = "สวัสดีค่ะ ยินดีต้อนรับสู่ระบบเสียงสังเคราะห์ภาษาไทย มะขาม"
phonemes = ipa_to_kokoro(g2p.convert(text))
token_ids = [0] + [cfg["vocab"][c] for c in phonemes if c in cfg["vocab"]] + [0]

# 4. Synthesize speech
audio = session.run(None, {
    "input_ids": np.array([token_ids], dtype=np.int64),
    "speed": np.array([1.0], dtype=np.float32),
})[0][0]

# 5. Save output audio (24 kHz)
sf.write("output.wav", audio, 24000)
print("Saved audio to output.wav")

---

Training & Dataset Curation

MakHam was trained using the verified pythainlp/MakHamTTS-dataset:

• Over 94,200 verified utterances (~100 hours of 24 kHz audio) curated from Thai Wikipedia, conversational logs, code-switching corpora, and Common Voice.
• Filtered through automated quality gates: duration trimming (>2.0s), SpeechBrain ECAPA-TDNN speaker verification, and Whisper CER consistency scoring.
• Kokoro base trained via two-stage StyleTTS2 with WavLM SLM adversarial prosody tuning.
• Distilled via the Paradee 1.1 framework into a unified TextStudent + VDecoder architecture.

---

Citation & Acknowledgements

If you use MakHam in your research or application, please cite:

@misc{makham2026,
  author       = {Wannaphong Phatthiyaphaibun},
  title        = {MakHamTTS},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/pythainlp/MakHamTTS}}
}

Credits

• Kokoro-82M: hexgrad/Kokoro-82M
• Paradee Framework: Sahil Mahendrakar, Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech Model, arXiv:2610.06817 (2026).
• FastThaiG2P: awslabs/FastThaiG2P
• PyThaiNLP: PyThaiNLP