Vietnamese Text to Speech & Voice Cloning (VieNeu-TTS v3 Turbo)

Powered by the open-source VieNeu-TTS model by Phạm Nguyễn Ngọc Bảo (@pnnbao97, Apache-2.0). Andrew Babo only packages and speed-optimises it for Apify — see Credits & source. No rental or usage fee: you pay Apify compute only.

Turn Vietnamese text into natural, studio-clean 48 kHz speech — with 23 built-in Vietnamese voices (North / Central / South accents, male & female), zero-shot voice cloning from a 3–8 second clip, English–Vietnamese code-switching, and inline emotion cues. Runs on CPU only, so it is cheap to run at scale on Apify.

Keywords: Vietnamese text to speech, Vietnamese TTS API, giọng đọc tiếng Việt AI, AI voice generator Vietnamese, voice cloning tiếng Việt, text to speech MP3, TTS for YouTube/TikTok voiceover.

What you can build

• YouTube, TikTok and Reels Vietnamese voiceovers
• Audiobooks and story narration (doc_truyen style voices)
• News reading and e-learning narration (tin_tuc style voices)
• IVR prompts, notifications and accessibility read-aloud
• Cloned brand/personal voice for repeatable content

Quick start

{
  "text": "Xin chào, đây là giọng đọc tiếng Việt tự nhiên do AI tạo ra.",
  "voice": "Adam",
  "format": "mp3"
}

Run it from the API:

curl -X POST \
  "https://api.apify.com/v2/acts/andrew_babo~vieneu-tts/runs?token=$APIFY_TOKEN&waitForFinish=180" \
  -H "Content-Type: application/json" \
  -d '{"text":"Xin chào, đây là giọng nói tiếng Việt.","voice":"Trúc Ly","format":"mp3"}'

The audio file lands in the run key-value store (audio.wav / audio.mp3 / audio.ogg) and one dataset row carries audio_url plus timing metrics.

Voice IDs (23 built-in Vietnamese voices)

Pass the exact name in the voice field (Vietnamese diacritics included). Default: Adam.

Voice ID	Gender	Accent (region)	Style
Adam (default)	Male	South (Nam)	Natural — tự nhiên
Phạm Tuyên	Male	North (Bắc)	Natural — tự nhiên
Xuân Vĩnh	Male	North (Bắc)	Natural — tự nhiên
Mạnh Dũng	Male	North (Bắc)	Natural — tự nhiên
Minh Quân	Male	North (Bắc)	Natural — tự nhiên
Quang Sơn	Male	Central (Trung)	Natural — tự nhiên
Minh Đức	Male	North (Bắc)	News — tin tức
Minh Triết	Male	South (Nam)	News — tin tức
Thanh Bình	Male	North (Bắc)	Storytelling — kể chuyện
Anh Khôi	Male	North (Bắc)	Storytelling — kể chuyện
Thái Sơn	Male	South (Nam)	Storytelling — kể chuyện
Đức Trí	Male	South (Nam)	Audiobook — đọc truyện
Trúc Ly	Female	North (Bắc)	Natural — tự nhiên
Đoan Trang	Female	North (Bắc)	Natural — tự nhiên
Ngọc Huyền	Female	North (Bắc)	Natural — tự nhiên
Ngọc Trân	Female	Central (Trung)	Natural — tự nhiên
Mai Anh	Female	North (Bắc)	News — tin tức
Thùy Dung	Female	South (Nam)	News — tin tức
Ngọc Linh	Female	North (Bắc)	Storytelling — kể chuyện
Thục Đoan	Female	South (Nam)	Storytelling — kể chuyện
Quỳnh Anh	Female	North (Bắc)	Audiobook — đọc truyện
Mỹ Duyên	Female	South (Nam)	Audiobook — đọc truyện
Kim Thanh	Female	South (Nam)	Audiobook — đọc truyện

Picking a voice:

• Ads / social videos: Adam, Trúc Ly, Thùy Dung
• News / corporate: Minh Đức, Mai Anh, Minh Triết
• Story & audiobook: Thái Sơn, Ngọc Linh, Quỳnh Anh, Kim Thanh
• Central accent: Quang Sơn, Ngọc Trân

If a voice name is misspelled the run fails with the full list of valid IDs.

Voice cloning (zero-shot, no training)

Give a public URL to a clean reference clip and the model speaks any Vietnamese text in that voice. Cloning overrides voice.

{
  "text": "Đây là giọng đã được nhân bản từ đoạn ghi âm mẫu.",
  "ref_audio_url": "https://example.com/reference.wav",
  "denoise": true,
  "format": "wav"
}

Reference clip checklist:

Requirement	Recommendation
Length	3–8 seconds (6 s is the sweet spot)
Content	One speaker, natural sentences, no music or background voices
Format	WAV or MP3, mono, 16 kHz or higher, no heavy compression
Delivery	Record in the emotion you want — emotion is transferred from the clip, not from a parameter
Hosting	Direct-download URL (S3, key-value store, CDN). Google Drive share links do not work.
Consent	Only clone voices you own or have written permission to use

denoise: true (default) cleans hiss and room noise before enrollment. Turn it off if your clip is already studio-quality and sounds over-processed after cloning.

Cloning tips:

• Same clip = consistent voice across many runs, so store it once and reuse the URL.
• A clip recorded cheerfully clones a cheerful voice; a calm clip clones a calm voice.
• Clips shorter than ~3 s or longer than ~10 s reduce similarity.

Emotion cues (Vietnamese, exactly 3 supported)

Write cues inline in the text:

Cue	Also accepted	Effect
[cười]	[cuoi], [laugh], [chuckle], [haha]	Laughter
[thở dài]	[tho dai], [sigh], [breath]	Sigh
[hắng giọng]	[hang giong], [clear throat], [cough]	Throat clear

Example:

Mình không tin nổi chuyện đó [cười]. Thôi kể tiếp nhé [hắng giọng].

Any other bracketed tag ([buồn], [khóc], [hét], …) is not an emotion. The actor strips unknown tags and reports them in the warnings field of the dataset row instead of reading them aloud.

Tone presets and inline directives

tone sets delivery style (it fills temperature / speed only when you did not set them yourself):

tone	Feel
tự nhiên	Natural (default)
trầm	Calm, slower
nhẹ nhàng	Gentle
kể chuyện	Storytelling
vui	Cheerful
hào hứng	Excited
trang trọng	Formal / news

Inline directives inside text (sticky until the next marker):

[tone: vui] Chào cả nhà, tin vui đây! [pause: 1s] [tone: trầm] Nhưng có một điều cần lưu ý.

[pause] inserts 0.5 s; [pause: 1s] inserts a custom silence (0.05–10 s).

Input reference

Field	Type	Default	Description
text	string	required	Vietnamese or Vi/En mixed text, up to 10,000 characters.
voice	enum	Adam	One of the 23 voice IDs above.
ref_audio_url	string	—	Reference clip URL for voice cloning; overrides voice.
denoise	boolean	true	Clean the reference clip before enrollment.
tone	enum	tự nhiên	Delivery preset (table above).
expressiveness	enum	medium	low / medium / high. Overridden by temperature.
temperature	number	0.8	0.1–1.5. Higher = more expressive, less stable.
speed	string	1.0	1.2 ≈ 20% faster.
pitch_preserve	boolean	true	Change speed without the chipmunk effect.
precision	enum	auto	fp32 (cleanest) / int8 (fastest) / auto follows quality.
quality	enum	smooth	smooth = seamless single pass for short text; fast = parallel, slightly faster.
single_pass	boolean	false	Force one engine to render the whole text.
trim_silence	boolean	auto	Trim edge silence; auto-off when emotion cues are present.
format	enum	wav	wav 48 kHz PCM, mp3 128 kbps, opus 64 kbps.

Output

{
  "audio_url": "https://api.apify.com/v2/key-value-stores/<store-id>/records/audio.mp3",
  "engine": "vieneu-v3-turbo",
  "chars": 49,
  "chunks": 1,
  "duration_s": 5.8,
  "load_ms": 7300.0,
  "synth_ms": 2900.0,
  "ms_total": 10300.0,
  "rtf": 0.5,
  "peak_ram_mb": 1100.0,
  "cores": 4,
  "sample_rate": 48000,
  "bytes": 556844,
  "format": "mp3"
}

rtf < 1 means synthesis is faster than realtime. warnings appears only when unknown bracketed tags were removed.

Standby mode (low-latency API)

Enable Standby to keep the model warm and skip the ~15–40 s cold model load:

curl -X POST "https://<actor-standby-url>/" \
  -H "Authorization: Bearer $APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"text":"Xin chào từ chế độ Standby.","voice":"Mai Anh","format":"mp3"}' \
  --output speech.mp3

The response body is the audio file; timing metrics come back in the X-TTS-Metrics header. Warm requests typically run 7–13 s for a short sentence on a 16 GB / 4 vCPU container.

Performance and cost notes

• CPU-only: no GPU required, 16 GB memory recommended (4 GB minimum).
• int8 is the fastest mode; fp32 is the reference-fidelity mode.
• Long narration: keep the whole passage in one run for prosody continuity, or split at sentence boundaries client-side and join with a short pause.

Limitations

• Vietnamese-first. English words inside Vietnamese sentences work; full English paragraphs are better served by an English TTS actor.
• Emotion cues are experimental model behaviour, not guaranteed on every sentence.
• Cloning quality depends heavily on the reference clip.
• Cold start pays the model load once per container.

Credits & source

The underlying model and Python library are not the work of the Actor publisher.

Original author: Phạm Nguyễn Ngọc Bảo — Speech AI Specialist and Medical NLP Researcher (Nhân dân Gia Định Hospital, Ho Chi Minh City, Vietnam), creator and maintainer of VieNeu-TTS.

Resource	Link
Author (GitHub)	https://github.com/pnnbao97
Source repository	https://github.com/pnnbao97/VieNeu-TTS
Models (Hugging Face)	https://huggingface.co/pnnbao-ump/VieNeu-TTS
Python package	https://pypi.org/project/vieneu/
Project website	https://www.vieneu.io

Model lineage: VieNeu-TTS is built on the neuphonic/neutts-air base and trained on the author's own Vietnamese speech datasets (pnnbao-ump/VieNeu-TTS-140h, then VieNeu-TTS-1000h); later v2/v3 lines add bilingual English–Vietnamese data and podcast-style multi-speaker support. This Actor runs the v3 Turbo line through the official vieneu package. Released under the Apache License 2.0.

What Andrew Babo did: Andrew Babo is not the model author. His role is to package VieNeu-TTS as an Apify Actor and optimise it for fast CPU inference on the Apify platform — a warm multi-engine pool, int8/fp32 precision selection, chunking and seam levelling for smooth joins, Standby warm start to cut cold-load time, inline tone / pause directives, emotion-cue validation, and MP3/Opus export.

Pricing: this Actor charges no rental and no usage fee. You pay only Apify platform compute for the time your run consumes.

Licensing and consent

VieNeu-TTS weights, preset voices and the vieneu library are distributed by the original author under the Apache License 2.0; the copyright stays with Phạm Nguyễn Ngọc Bảo. Keep that attribution if you redistribute anything derived from this Actor, and review the upstream model card and dataset terms before commercial distribution. For cloning, use only recordings you own or are authorized to process — do not impersonate anyone without permission.