SDK Overview

The vieneu package runs VieNeu-TTS v3 Turbo on your own machine. It defaults to the torch-free ONNX engine on CPU and switches to PyTorch on a CUDA GPU with no code change.

Where does this fit?

The SDK is the on-device path — free, open source (Apache 2.0), your hardware. If you would rather call a hosted API (including the proprietary v4 engine with higher cloning fidelity), see the Cloud API.

The pages after this one go deeper on one topic each:

• Install & backends — pip install vieneu, CPU vs GPU, precision, v3 Nano
• GPU batching — infer_batch, CUDA graphs, throughput numbers
• Streaming — infer_stream, concurrent streams
• Voice cloning — ref_audio, add_voice, denoise
• OpenAI-compatible server — /v1/audio/speech from the repo or Docker, plus the legacy v2 remote mode

Source

This section mirrors the Using the Python SDK part of the open-source README and is refreshed automatically (last sync 2026-09-16). If something here disagrees with the README, the README wins — open an issue there.

The vieneu SDK defaults to VieNeu-TTS v3 Turbo (48 kHz). The 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 — where inference is batched automatically (same API, no code change).

Quick Start

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

⚡On CPU the backbone runs fp32 by default (maximum fidelity). Need more speed? Pass Vieneu(precision="int8") — ~1.6× faster and ~4× smaller, but it requires a CPU with VNNI (AVX-512 VNNI / AVX-VNNI); on older CPUs int8 can produce garbled audio. precision only affects the CPU/ONNX path; on GPU it's ignored (PyTorch).

🪶 Still too slow, or deploying on a phone / ARM board? Try VieNeu-TTS v3 Nano (preview) — Vieneu(mode="v3nano"), ~3× faster than Turbo fp32 on CPU (RTF 0.11–0.22 on a desktop CPU), but noticeably lower quality (especially English / bilingual), 24 kHz, 11 preset voices + voice cloning. Details and caveats in the v3 Nano section below.

pip install vieneu

GPU (CUDA) — only if you have an NVIDIA GPU. On Linux pip install "vieneu[cuda]" is enough (PyPI torch ships CUDA there); on Windows install the CUDA torch first as below.

ℹ️ How fast is the GPU path? Since 3.7.0 every audio frame is one CUDA graph (acoustic decoder + sampling + repetition penalty + backbone step in a single replay — no torch.compile, no C++ toolchain needed). Measured on an RTX 3060: a 3.5 s sentence in 0.36 s; a 2-chunk paragraph (19 s) in 1.4 s; 16 chunks (154 s) in 2.8 s (RTF 0.02) — previously 2.3 s / 8.7 s / 16.7 s. The first call for each batch size pays ~0.5 s to capture the graph (kept afterwards; servers can call warm_fused() at start-up). VIENEU_FUSED_FRAME=0 restores the plain loop.

pip install torch==2.8.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128
pip install "transformers==4.57.6"   # pinned — most stable transformers for the GPU SDK
pip install vieneu

import time
from vieneu import Vieneu

# Default = v3 Turbo (48 kHz). GPU → PyTorch (auto-detected).
vieneu = Vieneu() # On a GPU machine you can still switch to ONNX/CPU if you prefer: Vieneu(backend="onnx")

# 1. Built-in voice by name — no reference clip needed
print("🔊 Generating speech...")

start_time = time.time()
audio = vieneu.infer("[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!", voice="Phạm Tuyên")
elapsed_time = time.time() - start_time

vieneu.save(audio, "output.wav")
print("✅ Saved to output.wav")

# Tính RTF (Real-Time Factor)
sample_rate = 48000
audio_duration = len(audio) / sample_rate
rtf = elapsed_time / audio_duration

print(f"\n⏱️  Thời gian xử lý: {elapsed_time:.3f}s")
print(f"🎵 Thời lượng audio: {audio_duration:.3f}s")
print(f"📊 RTF: {rtf:.4f}  ({'nhanh hơn' if rtf < 1 else 'chậm hơn'} real-time {1/rtf:.2f}x)" if rtf > 0 else "")

# List the built-in voices
voices = vieneu.list_preset_voices()
print(f"\n🎙️  {len(voices)} built-in voices available:")
for label, voice_id in voices:
    print(f"  - {label} ({voice_id})")

# 2. ⚡ Batch on GPU: infer_batch() runs many texts in ONE batched forward — same API.
#    On a CUDA GPU the chunks from every text share each forward step (big throughput
#    win). On CPU it still WORKS (no error) — just sequentially, so there's no batch
#    gain. Batch caps at max_batch_size (default 32; tune via Vieneu(max_batch_size=64)
#    or infer_batch(..., batch_size=64), or batch_size=1 to disable). A single long
#    infer() also auto-batches its own chunks. For real-time use, infer_stream() is the
#    streaming twin (GPU: 16 concurrent streams — see "Streaming" below). Uncomment to
#    try (GPU recommended):
#
# import time
# texts = [
#     "Chào cả nhà, hôm nay mình sẽ hướng dẫn các bạn cách cài đặt và sử dụng bộ giọng đọc mới.",
#     "Giọng nghe cực kỳ tự nhiên và truyền cảm, lại có thể chuyển đổi biểu cảm một cách linh hoạt.",
#     "Nếu thấy hữu ích, các bạn nhớ để lại một lượt thích và chia sẻ video này cho mọi người nhé!",
# ] * 10   # 30 texts — enough to fill the batch and really show the GPU throughput win
# t0 = time.time()
# audios = vieneu.infer_batch(texts, voice="Minh Quân Pro")
# elapsed = time.time() - t0
# total_audio = sum(len(a) for a in audios) / 48_000
# print(f"⚡ {len(texts)} texts | audio {total_audio:.1f}s | wall {elapsed:.1f}s | RTF {elapsed/total_audio:.3f}")
# for i, a in enumerate(audios):
#     vieneu.save(a, f"batch_{i}.wav")

Streaming (real-time) 🔊

v3 Turbo streams frame by frame on both backends. GPU (PyTorch): first audio in ~115 ms and 16 concurrent streams on one RTX 3060 (continuous batching — one CUDA graph serves every infer_stream call, each keeping RTF ≈ 0.5–0.6). CPU (ONNX): first audio in ~140 ms (int8) / ~300 ms (fp32), one stream (two with int8). Just iterate infer_stream:

from vieneu import Vieneu
vieneu = Vieneu()                                  # GPU → PyTorch + stream scheduler; no GPU → ONNX/CPU
for chunk in vieneu.infer_stream("Xin chào các bạn!", voice="Mai Anh"):
    play(chunk)                                   # np.float32 @ 48 kHz — play/write as it arrives

Calling infer_stream from many threads at once is the intended way to serve many listeners on a GPU (Vieneu(max_streams=16) sets the ceiling).

An OpenAI-compatible streaming API (POST /v1/audio/speech, pcm/wav, chunked or SSE — works with the OpenAI SDK, Pipecat, LiveKit, …) is in apps/openai_speech.py:

# Pick ONE of these — all serve http://localhost:8000/v1/audio/speech
uv run python -m apps.openai_speech                                  # from the repo (auto-detects GPU/CPU)
docker compose -f docker/docker-compose.yml --profile api-gpu up     # or: Docker, GPU
docker compose -f docker/docker-compose.yml --profile api-cpu up     # or: Docker, CPU only

📊 docs/streaming.md — every measurement on an RTX 3060 (TTFA / RTF / streams vs max_streams), estimates for smaller GPUs, and the CPU numbers. The older browser demo is still at apps/web_stream.py.

Available Voices

The v3 Turbo engine includes 25 preset voices covering 3 regions (North, Central, South) with diverse genders and speaking characters. list_preset_voices() (and the Web UI / API voice lists) show them in this order:

• ⭐ Editors' picks — the 10 we recommend starting with, hand-selected for naturalness and stability: Adam bựa, Trúc Ly, Anh Khôi, Mai Anh, Minh Quân Pro (default; "Minh Quân" still works as an alias), Thùy Dung, Thiền Tâm Đức, Ngọc Huyền, Quang Sơn, Ngọc Trân
• Northern (Bắc): Minh Đức, Phạm Tuyên, Xuân Vĩnh, Thanh Bình, Ngọc Linh, Đoan Trang, Quỳnh Anh, Mạnh Dũng (+ picks above)
• Central (Trung): Quang Sơn, Ngọc Trân
• Southern (Nam): Adam, Thái Sơn, Thục Đoan, Minh Triết, Mỹ Duyên, Đức Trí, Kim Thanh (+ Thùy Dung)

Reading style — deprecated ⚠️

warning

style is deprecated on v3 Turbo and has no effect. The reading style is already baked into the reference itself (the speaker embedding + reference codes of the preset voice or of your cloned clip), so every generation follows the reference and comes out in its natural reading style.

The style argument is still accepted by infer, infer_stream, infer_batch and add_voice so existing code keeps running — whatever you pass ("tin_tuc", "doc_truyen", …) is simply ignored. New code should just omit it.

# Old code — still runs, but `style` is ignored
audio = vieneu.infer("Bản tin sáng nay.", voice="Minh Quân Pro", style="tin_tuc")

# New code — pick the reading character through the voice / reference clip instead
audio = vieneu.infer("Bản tin sáng nay.", voice="Minh Quân Pro")

Emotion cues (experimental)

Inline tags are supported anywhere in the text: [cười] (chuckle), [thở dài] (sigh), [hắng giọng] (clear throat).

audio = vieneu.infer("Nghe hay quá đi [cười]. Để mình nói tiếp [hắng giọng].", voice="Minh Quân Pro")

Voice cloning

Clone any voice from a short reference clip. The clip is cleaned up automatically (background noise removed, and trimmed to ≤ 8 seconds) before cloning — keep denoise=True unless your clip is already clean.

audio = vieneu.infer(
    "Đây là giọng được nhân bản tức thì.",
    ref_audio="my_voice.wav",   # a 3–8s reference clip
    denoise=True,               # default; set False if the clip is already clean
)
vieneu.save(audio, "cloned.wav")

Save & reuse a cloned voice

Register a reference once with add_voice, then use it by name like a built-in voice.

# Enroll a voice (denoises + extracts the speaker profile once)
vieneu.add_voice("Giọng của tôi", "my_voice.wav")

# Now reuse it anywhere, including the conversation mode
audio = vieneu.infer("Câu này dùng giọng đã lưu.", voice="Giọng của tôi")

# Persist your voices so they load next session
vieneu.save_voices()                 # writes to the default voices file
# vieneu.remove_voice("Giọng của tôi")

# Add a voice you already cleaned yourself → skip denoising
vieneu.add_voice("Giọng sạch", "already_clean.wav", denoise=False)

Clean up a clip on its own

Get the denoised audio without synthesizing anything (e.g. to inspect or store it):

wav, sr = vieneu.denoise("noisy.wav", out_path="clean.wav")   # 44.1 kHz mono

Note: denoise, add_voice, and voice cloning work on every backend — the torch-free CPU/ONNX install included (the whole cloning pipeline runs on onnxruntime + soxr + kaldi-native-fbank). v3 Nano below clones the same way (its cloning graphs are fetched on first use).

v3 Nano (preview) — for edge devices / weak CPUs only 🪶

warning

v3 Turbo remains the default and the recommended model. Use v3 Nano only when Turbo is too slow on your hardware (old laptops, mini PCs, ARM boards, CPUs without AVX-512/VNNI where the int8 Turbo build produces garbled audio). Nano is a 48M-parameter flow-matching model (ONNX, CPU, torch-free) and it trades quality for speed:

• Lower quality than v3 Turbo — most noticeably on English and code-switched (En-Vi) text. Vietnamese is close; English words come out with a Vietnamese accent and are less stable.
• 24 kHz output (Turbo: 48 kHz).
• 11 preset voices + voice cloning (ref_audio, add_voice, encode_reference work like Turbo; the three cloning graphs, ~110 MB, download on first use).
• No frame-level streaming — infer_stream yields one finished chunk at a time.

Measured on the same desktop CPU (12th-gen Intel i7, 6 ONNX Runtime threads, ~9 s of speech):

Engine	RTF ↓	Sample rate	Load time
v3 Turbo ONNX fp32 (default on CPU)	0.62	48 kHz	~19 s
v3 Turbo ONNX int8	0.37	48 kHz	~14 s
v3 Nano, 16 steps, cfg 3 (default)	0.22	24 kHz	~3 s
v3 Nano, 8 steps, sway −1	0.11	24 kHz	~3 s

RTF = compute time ÷ audio duration (lower is faster; 0.22 = 4.5× faster than real time). The ratio carries over to slower machines: expect Nano to be roughly 1.7× faster than Turbo int8 and ~3× faster than Turbo fp32, with a 282 MB download instead of Turbo's.

from vieneu import Vieneu

tts = Vieneu(mode="v3nano")                      # ONNX, CPU, torch-free
audio = tts.infer("Xin chào, mình là giọng đọc của VieNeu Nano.", voice="Minh Quân")
tts.save(audio, "nano.wav")                      # 24 kHz

tts.list_preset_voices()                         # Adam, Ái Hân, Mỹ Duyên, Đức Trí, Hữu Quân, Xuân Tiên, Mai Anh, Trúc Ly, Anh Khôi, Minh Quân, Mạnh Dũng
audio = tts.infer("Bản nhanh cho máy rất yếu.", voice="Ái Hân", steps=8, sway=-1)   # ~2× faster

Knobs: steps (Euler steps, 16 default; 8 ≈ 2× faster, slightly rougher — pair with sway=-1), cfg (classifier-free guidance, 3.0 default; cfg=0 halves compute but hurts intelligibility), speed, seed, threads. Emotion cues [cười] [thở dài] [hắng giọng] work as on Turbo.