🦜 VieNeu-TTS v3 Turbo

Overview

VieNeu-TTS v3 Turbo is the next generation of Vietnamese TTS — 48 kHz high-fidelity speech, 23 built-in preset voices across three regions (North / Central / South), instant voice cloning, real-time streaming with an OpenAI-compatible API (16 concurrent streams on one RTX 3060), inline emotion cues, and seamless bilingual (En–Vi) code-switching.

The reference implementation is the vieneu Python SDK (v3.7.1). Its minimal install is torch-free: on CPU everything runs on ONNX Runtime (PyTorch is never imported), and on a CUDA machine it auto-switches to the PyTorch engine with automatic batching and a continuous-batching stream scheduler — same API, no code change.

[!IMPORTANT] What's new in SDK v3.7:

• GPU streaming with continuous batching — infer_stream on CUDA runs on one CUDA graph shared by every caller: first audio in ~115 ms, 16 concurrent streams under 200 ms on a single RTX 3060 (32 max).
• OpenAI-compatible streaming API + Docker — POST /v1/audio/speech (pcm/wav, chunked or SSE) works with the OpenAI SDK, Pipecat, LiveKit; docker compose --profile api-gpu / api-cpu. Numbers, sizing and CPU-only notes: docs/streaming.md.
• 23 preset voices covering North / Central / South, both genders and several reading characters (default: Minh Quân).
• LoRA fine-tuning of v3 Turbo on one consumer GPU.
• Since 3.3: torch-free voice cloning on CPU, optional int8 CPU backbone (Vieneu(precision="int8"), ~2× faster; fp32 is the default), sliding-window repetition penalty.

🏗️ Architecture & Credits

The VieNeu-TTS v3 Turbo architecture is an original design by the author, Phạm Nguyễn Ngọc Bảo, and is trained from scratch on ~10,000 hours of English–Vietnamese speech — it is not a fine-tune, distillation, or adaptation of any existing TTS model.

• Model architecture & training: designed and trained from scratch by Phạm Nguyễn Ngọc Bảo — https://github.com/pnnbao97
• Audio codec: MOSS-Audio-Tokenizer-Nano (OpenMOSS-Team) — 48 kHz neural audio codec.
• Phonemizer: sea-g2p — fast Vietnamese/English grapheme-to-phoneme, also by the author.

Tác giả: Phạm Nguyễn Ngọc Bảo

---

🔥 Quick Start (Web UI)

git clone https://github.com/pnnbao97/VieNeu-TTS.git
cd VieNeu-TTS

• Option 1: CPU & macOS (minimal, torch-free) — recommended — runs v3 Turbo via ONNX

  uv sync
  

  ⚡ Use uv sync, not pip install, for the fastest CPU inference — it reproduces the locked environment with the optimized ONNX Runtime build. On Apple Silicon this ONNX/CPU path is faster than the MPS/PyTorch build.

• Option 2: GPU (CUDA ≥ 12.8) — v3 Turbo on GPU (PyTorch): batched generation + multi-stream real-time streaming

  uv sync --extra cuda
  

Start the Web UI:

uv run vieneu-web

The UI opens at http://127.0.0.1:7860 with a Default voice tab, a Voice Cloning tab, and a Conversation tab (batched multi-speaker podcasts).

---

📦 Using the Python SDK (vieneu)

CPU (default) — torch-free, runs v3 Turbo via ONNX Runtime. Most users want this:

pip install vieneu

GPU (CUDA) — only if you have an NVIDIA GPU; install a CUDA build of PyTorch yourself first:

pip install torch==2.8.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128
pip install "transformers==4.57.6"   # Qwen3 backbone + MOSS codec (pinned — most stable)
pip install vieneu

ℹ️ When is GPU actually worth it? Two cases: bulk / long-form synthesis (many chunks share one batched forward) and serving many listeners (the stream scheduler batches every infer_stream call — 16 concurrent real-time streams on an RTX 3060). For a single short, interactive call the torch-free CPU/ONNX path is fast enough (first audio ~140 ms with int8, ~300 ms with fp32) and needs no GPU.

Full features guide

from vieneu import Vieneu
from time import time

# Default = v3 Turbo (48 kHz). CPU → ONNX (torch-free, fp32); GPU → PyTorch (auto-detected).
tts = Vieneu()                    # fp32 backbone (default, max fidelity)
# tts = Vieneu(precision="int8")  # CPU only: ~2x faster, needs a CPU with VNNI

text = """[cười] Trời ơi, cái giọng nó tự nhiên mà nó mượt mà dã man, nghe không khác gì người thật luôn. Giờ thì tha hồ mà quẩy content với cả kho giọng nói đa dạng, đủ mọi sắc thái biểu cảm. Mọi người bật loa lên rồi cùng trải nghiệm thử với mình nhé!"""

# 1. Default voice (Minh Quân) — 48 kHz, no reference needed
start = time()
audio = tts.infer(text)
tts.save(audio, "output.wav")
print(f"Time taken: {time() - start:.2f} seconds")

# 2. Built-in voices by name
for label, voice_id in tts.list_preset_voices():
    print(label, voice_id)
audio = tts.infer("Mình là Xuân Vĩnh nè!", voice="Xuân Vĩnh")
tts.save(audio, "output_xuan_vinh.wav")

# 3. Emotion / non-verbal cues — EXPERIMENTAL: [cười] [thở dài] [hắng giọng]
audio = tts.infer("Nghe hay quá đi [cười]. Để mình nói tiếp [hắng giọng].", voice="Phạm Tuyên")

# 4. Instant voice cloning from a 3–8s reference clip (works on the torch-free CPU install too)
audio = tts.infer("Đây là giọng được nhân bản tức thì.", ref_audio="my_voice.wav", denoise=True)

[!TIP] A temperature around 0.8 gives the most stable result for v3 Turbo. Higher values add expressiveness but can be less stable.

🔊 Real-time streaming

v3 Turbo streams frame by frame on both backends. GPU: first audio in ~115 ms, and every infer_stream call shares one CUDA graph (continuous batching) — 16 concurrent streams stay real-time on a single RTX 3060 (RTF ≈ 0.5–0.6, 32 streams max). CPU (ONNX): first audio ~140 ms (int8) / ~300 ms (fp32), RTF < 1, one stream (two with int8).

vieneu = Vieneu()   # GPU → PyTorch + stream scheduler; no GPU → ONNX/CPU
for chunk in vieneu.infer_stream("Xin chào! Đây là chế độ streaming của VieNeu, phát tới đâu nghe tới đó.", voice="Mai Anh"):
    play(chunk)   # np.float32 @ 48 kHz, play/write as it arrives

OpenAI-compatible API — POST /v1/audio/speech (pcm/wav, chunked body or SSE), drop-in for the OpenAI SDK, Pipecat, LiveKit:

uv run python -m apps.openai_speech                                  # → http://localhost:8000
# (or: docker compose -f docker/docker-compose.yml --profile api-gpu up)

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="x")
with client.audio.speech.with_streaming_response.create(
    model="vieneu-v3-turbo", voice="Mai Anh", response_format="pcm",
    input="Xin chào! Đây là chế độ streaming của VieNeu, phát tới đâu nghe tới đó.",
) as r:
    for chunk in r.iter_bytes(4096):   # s16le 48 kHz mono, as it is generated
        play(chunk)

All measurements (TTFA / RTF / streams vs max_streams on an RTX 3060), estimates for smaller GPUs and the CPU-only numbers: docs/streaming.md.

⚡ Batched generation (GPU)

infer_batch() runs many texts in one batched forward — same API on every backend (on CPU it still works, just sequentially). The batch caps at max_batch_size (default 32); pass batch_size=1 to disable. A single long infer() also auto-batches its own chunks.

audios = vieneu.infer_batch(texts, voice="Mai Anh")   # or infer_batch(..., batch_size=64)

🦜 Voice cloning & saved voices

# Clone from a 3–8s clip; the reference is auto-denoised and trimmed to ≤ 8s
audio = vieneu.infer("Chào bạn, đây là giọng của tôi.", ref_audio="voice.wav", denoise=True)

# Enroll once, then reuse by name like a built-in voice
vieneu.add_voice("Giọng của tôi", "voice.wav")
audio = vieneu.infer("Câu này dùng giọng đã lưu.", voice="Giọng của tôi")

# Just clean up a clip (no synthesis)
wav, sr = vieneu.denoise("noisy.wav", out_path="clean.wav")

denoise, add_voice and cloning work on every backend, including the torch-free CPU/ONNX install.

⚠️ Reading style is deprecated

style is still accepted by infer, infer_stream, infer_batch and add_voice so existing code keeps running, but it is ignored on v3 Turbo: the reading style is already baked into the reference itself (the speaker embedding + reference codes of the preset voice or of your cloned clip). Pick the reading character through the voice instead.

---

🎭 Preset Voices (23)

Call any of them by name via voice="<name>" — no reference audio required. Default voice: Minh Quân. (Generated from the SDK's voices_v3_turbo.json; vieneu.list_preset_voices() is authoritative.)

For any other voice, use voice cloning with a short reference clip (ref_audio="...").

---

🔬 Model Variants

---

📜 Usage Rights & Licensing FAQ

Does Apache-2.0 cover every artifact in this repository? Yes. The license applies to all artifacts shipped here — model.safetensors, the ONNX exports, configs and tokenizers, and the bundled preset-voice assets (speaker embeddings + reference codes in voices_v3_turbo.json).

May I use the preset voices and the generated audio commercially? Yes. The bundled preset voices are distributed under the same Apache-2.0 license as the rest of the repository, and audio generated with them may be used in commercial and monetized content (voice-over, videos, products, services) — no additional license or fee.

Did the speakers behind the preset voices consent to AI training and synthetic speech? Yes. The speakers (or rightsholders) behind the shipped preset-voice assets granted appropriate rights and consent for their voice data to be used in AI training and synthetic speech generation, which is what allows those assets to be distributed under Apache-2.0 for both non-commercial and commercial synthetic audio generation.

What about the training dataset? The detailed internal data-collection and processing pipeline for the training corpus is not publicly disclosed, and the VieNeu-TTS-10k-ENVI dataset is gated. The confirmations above cover the preset voices shipped in this repository and the model weights released here, which are the artifacts you actually redistribute or generate audio with.

Which preset list is authoritative? vieneu.list_preset_voices() at the version you have installed. This card documents SDK v3.7.1 (23 voices, default Minh Quân); earlier revisions shipped fewer voices under partly different names (3.3.0: 20 voices, default Adam), so pin the SDK version if the exact roster matters to you.

Third-party components — all permissively licensed, keep their notices when redistributing:

• MOSS-Audio-Tokenizer-Nano (OpenMOSS-Team) — Apache-2.0.
• sea-g2p — phonemizer, by the same author as this project.

[!WARNING] Voice cloning is your responsibility. The consent confirmation above covers the bundled preset voices only. If you clone a voice from your own reference clip, you must have the right to use that person's voice. Do not clone real people without their permission, and do not use this model to impersonate, defraud, or produce misleading content.

---

License

This model package is distributed under Apache License 2.0, matching the upstream model repository.

When you reuse, redistribute, or convert these assets, please keep the license notice and attribution intact for both:

• the original project: pnnbao97/VieNeu-TTS
• this Hugging Face package: pnnbao-ump/VieNeu-TTS-v3-Turbo

If you bundle additional third-party assets, their own licenses still apply as well.

---

📑 Citation

@misc{vieneutts2026,
  title        = {VieNeu-TTS v3 Turbo: 48kHz Vietnamese Text-to-Speech with Instant Voice Cloning and Emotion Control},
  author       = {Pham Nguyen Ngoc Bao},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/pnnbao-ump/VieNeu-TTS-v3-Turbo}}
}

---

Made with ❤️ for the Vietnamese TTS community