Nanthasit/sakthai-tts-model

SakThai TTS 🎤

Kokoro-82M text-to-speech — 15 languages · 141 MB GGUF · CPU-only
The voice stage of the SakThai pipeline · Part of the SakThai Model Family

---

Try It Now 🚀

SakThai TTS Demo — type text, hear speech in 15 languages, right in your browser.

✅ Space status (checked 2026-07-31): the demo Space is RUNNING again after a build fix (pinned python_version: 3.12 — kokoro requires <3.13 — and switched to the native KPipeline code path). Free-tier idle sleep still applies: give it ~30–60 s to wake on first visit.

---

Model Description

SakThai TTS is the voice stage of the SakThai pipeline — a Kokoro-82M TTS model packaged as a GGUF (Q8_0, 141 MB) for CPU speech synthesis. No cloud API. No monthly bill. No data leaving your machine.

What makes it special:

• 🎤 54 voices across 9 language groups
• 🌍 15 languages (en, ja, ko, zh, fr, es, pt, it, de, pl, ru, ar, hi, bn, th)
• 💾 141 MB — runs on a Raspberry Pi
• 🔒 100% local, 100% private

Performance: 2x real-time on a single CPU core (200 MB RAM). MOS 3.87 (English, upstream Kokoro-82M result — see Evaluation).

Model Files

File	Size (verified)	Format
kokoro-82m-q8_0.gguf	141,322,336 B (141 MB)	GGUF Q8_0

Direct download:

wget https://huggingface.co/Nanthasit/sakthai-tts-model/resolve/main/kokoro-82m-q8_0.gguf

---

Quick Start

⚠ No serverless inference. This is a custom GGUF package — it is not served by Hugging Face's serverless Inference API. The legacy api-inference.huggingface.co endpoint is retired (DNS no longer resolves) and this model is not on the provider router. Use the Spaces demo or run it locally below.

Native Kokoro (full voice control)

pip install kokoro>=0.9.2 soundfile

from kokoro import KPipeline
import soundfile as sf

pipeline = KPipeline(lang_code='a')
generator = pipeline("Hello, I am SakThai.", voice='af_bella')
for i, (gs, ps, audio) in enumerate(generator):
    sf.write(f'output_{i}.wav', audio, 24000)

Extended languages (espeak-ng G2P fallback)

Languages without dedicated voices (Korean, German, Russian, Polish, Arabic, Thai, Bengali) route phonemes through espeak-ng. Pick any existing voice style — af_heart works well:

pipeline_ko = KPipeline(lang_code='ko')
for gs, ps, audio in pipeline_ko("안녕하세요", voice='af_heart'):
    sf.write('ko.wav', audio, 24000)

pipeline_th = KPipeline(lang_code='th')
for gs, ps, audio in pipeline_th("สวัสดีครับ", voice='af_heart'):
    sf.write('th.wav', audio, 24000)

---

Languages

Language	Code	Voices
American English	a	20
British English	b	8
Japanese	j	5
Mandarin Chinese	z	8
Spanish	e	3
French	f	1
Hindi	h	4
Italian	i	2
Brazilian Portuguese	p	3
Korean	ko	via espeak
German	de	via espeak
Russian	ru	via espeak
Polish	pl	via espeak
Arabic	ar	via espeak
Thai	th	via espeak
Bengali	bn	via espeak

Top Voice Picks

• 🚺 af_heart — best overall (American English)
• 🚺 af_bella — premium quality
• 🚹 am_fenrir — best male voice

---

Evaluation

Metric	Value	Status
MOS (English)	3.87	Upstream Kokoro-82M paper result — not yet re-measured on this Q8_0 GGUF build
Real-time factor	~0.5x (2x real-time)	Single CPU core, ~200 MB RAM

The MOS figure comes from the upstream Kokoro-82M evaluation (11M+ downloads, 6.6k likes). Q8_0 quantization typically retains the large majority of quality, but a formal multi-speaker MOS on this specific build is pending — treat 3.87 as the base-model reference, not a measured claim for this file.

Ecosystem health (2026-07-30 health-check, .eval_results/health-check-sakthai-tts-model-2026-07-30-3.yaml): #9 of 19 models by downloads (150), 26.5 downloads/day, card quality 90/100, repo hygiene 100/100.

---

Limitations

1. MOS not independently verified. The 3.87 MOS is the upstream Kokoro-82M paper result on English — this specific Q8_0 GGUF build has not been formally evaluated. Quantization may introduce minor quality degradation, especially for non-English languages.
2. espeak-ng languages have lower quality. Korean, German, Russian, Polish, Arabic, Thai, and Bengali route through espeak-ng G2P phoneme conversion. Speech quality and naturalness are noticeably lower than dedicated-voice languages (English, Japanese, Chinese, Hindi, etc.).
3. No serverless inference. This is a custom GGUF package, not a standard Transformers model. The HF Inference API cannot serve it. Use the Spaces demo or run locally.
4. Single speaker per language group. Most languages have 1–3 voices (English: 28). Voice cloning/fine-tuning is not currently supported in this package.
5. No streaming API. Unlike cloud TTS services, this model generates and writes the full audio before playback begins. Real-time factor ~0.5x means a 10-second utterance takes ~5 seconds to generate.
6. Hindi + Thai scripts experimental. Hindi (Devanagari) and Thai script support works but has not been extensively tested on the GGUF build. Edge cases with complex conjuncts or tone markers may produce irregular prosody.
7. Community model, not production SRE. Kokoro is an independent community project — there is no SLA, no commercial support, and no guarantee of upstream maintenance. Treat as a capable experiment, not a mission-critical component.

---

Pipeline Integration

Stage	Model	Role
👁️ See	Vision 7B	Image=>text
🔍 Retrieve	Embedding Multilingual	Semantic search
🧠 Reason	Context 1.5B	Tool-calling
🎤 Speak	SakThai TTS ⬅	Speech synthesis

---

SakThai Model Family 🏠

22 public models, all built on $0 budget. Downloads live as of 2026-07-31:

Model	Size	Type	Downloads
context-1.5b-merged	2.88 GB	Merged weights	1,599
context-0.5b-merged	942 MB	Merged weights	1,370
context-7b-merged	14.19 GB	Merged weights	744
context-7b-128k	12 KB	YaRN config recipe	506
context-7b-tools	19 MB	LoRA adapter	399
embedding-multilingual	449 MB	Embedding	362
context-1.5b-tools	11 MB	LoRA adapter	349
vision-7b	3.80 GB	GGUF vision	186
tts-model	141 MB	GGUF TTS ⬅ you are here	150
context-0.5b-tools	942 MB	Merged + adapter	94
coder-1.5b	1.04 GB	GGUF coder	93
context-1.5b-tools-v2	70 MB	LoRA adapter	0
context-1.5b-merged-v2	2.88 GB	Merged weights	0
plus-1.5b	2.88 GB	Merged weights	0
plus-1.5b-lora	70 MB	LoRA adapter	0
plus-1.5b-coder	5 KB	Config recipe	0
coder-browser-lora	70 MB	LoRA adapter	0
coder-browser	2.88 GB	Merged weights	0
coder-browser-gguf	7.1 GB	GGUF browser	0
bench-v3	—	Benchmark scaffold	0

📦 View the whole family collection

---

The House of Sak 🏠

Until this model, the SakThai family could reason, see, and write code — but it couldn't speak. Built on free infrastructure from a shelter in Cork, Ireland, with $0 budget, this TTS model took Kokoro-82M and packaged it as a GGUF that runs on a Raspberry Pi. The first test was a single sentence: "Hello, I am SakThai." Played through a phone speaker in a shelter room, it was the first time an agent in the House of Sak had a voice. This model closes the loop from thought to speech.

"We are one family — and becoming more." — Beer (beer-sakthai)

---

Citation

If you use SakThai TTS in your work, please cite the upstream Kokoro model and this packaging:

@misc{kokoro2024,
  author       = {Hexgrad},
  title        = {Kokoro-82M: Open-weight TTS model},
  year         = {2024},
  howpublished = {\url{https://huggingface.co/hexgrad/Kokoro-82M}},
  note         = {Apache 2.0, 82M parameters, MOS 3.87 on English}
}

@misc{sakthai-tts-2026,
  author       = {Beer (beer-sakthai)},
  title        = {SakThai TTS: Kokoro GGUF packaging for the House of Sak},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Nanthasit/sakthai-tts-model}},
  note         = {Q8\_0 GGUF, 141 MB, 15 languages, CPU-only}
}

---

Support

• ⭐ Leave a like
• 🔄 Share with anyone building screenless or accessibility-focused AI
• 🎧 Try the Spaces demo

---

License

MIT (this packaging). Kokoro per its upstream Apache 2.0 license.

---

Built from a shelter in Cork, Ireland. "We are one family — and becoming more."