You can deploy this model and start earning money today!

trungnguyentran/
vietts-gguf

Name: vietts-gguf
Availability: InStock

Views

No views yet

VieTTS GGUF — Vietnamese Text-to-Speech (Local, GPU-accelerated)

Vietnamese TTS with instant voice cloning. Runs locally on Apple Silicon (Metal) or NVIDIA GPUs (CUDA) via llama-cpp-python. No cloud API needed.

38 preset voices (19 male, 19 female) with regional variety
Voice cloning from 3-10s reference audio
Long-form synthesis (minutes-scale) with voice consistency
~80 tok/s on Apple M4 (1.6× realtime)

Quick Start

bash

# Install (macOS with Metal)
CMAKE_ARGS="-DGGML_METAL=on" pip install llama-cpp-python
pip install soundfile sea-g2p onnxruntime transformers

# Single sentence
python infer.py --voice nam_long \
    --text "Xin chào, tôi là trợ lý AI của bạn."

# Long-form (2 minutes)
python infer.py --voice anh --text-file story.txt \
    --target-seconds 120 --out story.wav

# With speaker guard (prevents voice drift in long audio)
pip install resemblyzer librosa
python infer.py --voice nam_long --text-file story.txt \
    --target-seconds 120 --speaker-guard --out story.wav

# List all voices
python infer.py --list-voices --voice dummy

Recommended Parameters

Parameter	Value	Effect
`--temperature`	0.55	Lower = more consistent voice, less expressive
`--top-k`	50	Token sampling diversity
`--max-new-tokens`	900	~18s max per chunk (model's trained limit)
`--context-seconds`	6	First chunk anchors subsequent chunks
`--spk-threshold`	0.70	Speaker guard: reject drifted chunks
`--lock-threshold`	0.72	Voice lock: consistency between chunks
`--retries`	4	Extra attempts when speaker guard rejects

Voice Catalog

Male Voices

Voice ID	Region	Style	Ref (s)
nam_hung	—	deep, low, narrative	6.0
nam_duc	north	clear, moderate, narrative	5.9
nam_khoa	south	bright, high, conversational	5.3
nam_minh	south	clear, moderate, conversational	6.0
nam_long	south	warm, steady, conversational	6.0
nam_khuong	south	clear, moderate	6.0
nam_son	north	deep, low, narrative	6.0
nam_hieu	—	bright, high	6.0
nam_tuan	south	warm, steady	6.0
nam_vui	south	clear, moderate, conversational	6.0
anh	south	warm, calm, deep	6.0
nhat_phong	north	warm, measured, professional	6.0
uc	south	calm, steady, narrative	6.0
trung	south	deep, calm, steady	3.7
trung_caha	south	clear, firm, informative	3.1
trieu_duong	south	deep, low, narrative	3.3
tung	south	deep, steady, authoritative	5.0
ninh	neutral	deep, warm, resonant	4.7
quang	south	bright, energetic, young	3.3

Female Voices

Voice ID	Region	Style	Ref (s)
nu_ha	north	clear, moderate, narrative	6.0
nu_linh	south	high, energetic, conversational	6.0
nu_phuong	—	clear, moderate	6.0
nu_hoa	north	warm, low, narrative	6.0
nu_lan	north	clear, moderate, narrative	6.0
nu_my	north	high, energetic	4.7
nu_ngoc	south	bright, expressive, conversational	6.0
nu_thanh	—	clear, moderate, narrative	6.0
nu_dieu	north	bright, expressive, narrative	6.0
nu_thu	north	warm, low, narrative	6.0
trang	neutral	clear, smooth, direct	3.4
tham	neutral	soft, intimate, gentle	6.0
nhu	north	soft, gentle, warm	6.0
mai	north	bright, cheerful, expressive	6.0
ngan	neutral	smooth, professional, clear	6.0
thao	north	clear, bright, energetic	6.0
hien	north	warm, steady, natural	6.0
kenh	north	bright, youthful, dynamic	4.3
huyen	north	warm, expressive, conversational	5.4

How It Works

Text is phonemized with sea-g2p (Vietnamese IPA)
Phonemes + reference voice codes form the LM prompt
Qwen3-0.6B (GGUF Q8_0) generates NeuCodec speech tokens
NeuCodec ONNX decoder converts tokens to 24kHz audio

For long-form: text is chunked by sentence, first chunk's audio is used as a stable voice anchor for all subsequent chunks. Optional speaker guard (resemblyzer d-vector) rejects chunks that drift from the reference voice.

Model Details

Backbone: Qwen3-0.6B (extended with 65,536 speech tokens)
Codec: NeuCodec FSQ, single codebook, 50 tokens/second
Quantization: GGUF Q8_0 (~712MB)
Training data: viVoice (~1,017h Vietnamese, 630+ speakers)
Output: 24kHz mono WAV

Files

vietts-qwen3-0.6b-vi-ext-q8_0.gguf   # LM weights (Q8_0)
tokenizer.json                         # Tokenizer
tokenizer_config.json
added_tokens.json
special_tokens_map.json
voices.json                            # 38 preset voices with codes
infer.py                               # Inference script
README.md

MLX Q4 (Fastest on Apple Silicon)

The mlx-q4/ subfolder contains a 4-bit MLX-quantized model (~395 MB) that runs at ~95 tok/s (1.9x realtime) on Apple M4 -- 35% faster than GGUF Q8 with equal or better audio quality.

Model	Size	Speed (M4)	Quality
GGUF Q8_0	679 MB	~80 tok/s (1.6x)	Baseline
MLX Q4	395 MB	~95 tok/s (1.9x)	Equal or better

License

Model weights: research use. Voice presets derived from viVoice (CC-BY-NC-SA-4.0) and ElevenLabs preview clips (subject to their terms).

trungnguyentran/vietts-gguf