Breaking

ValtecTTS Launches: Lightweight Vietnamese TTS With CPU Voice Cloning

July 5, 2026 at 01:39 EDT

Media Generation
Open Source

A lightweight Vietnamese-focused text-to-speech model, "ValtecTTS," has been released and updated on GitHub. With 74.8M parameters and GPU-free CPU operation, it supports zero-shot voice cloning that reproduces a speaker's voice from just 3–10 seconds of reference audio.

valtecAI-team · Open Source Release

Valtec Vietnamese TTS: voice cloning that runs on a CPU

A lightweight 74.8M-parameter speech model built for tonal, low-resource Vietnamese — cloning any voice from 3–10 seconds of audio at 3–4× real-time speed, no GPU required.

74.8M

parameters (zero-shot) ≈ 285MB in FP32

3–10s

reference audio needed to clone a voice

CPU

only — no GPU required to generate

Faster than real-time — and better on longer text

Real-time factor (audio length ÷ generation time). Taller column = more seconds of speech produced per second of compute.

2.1×

Short
16 chars · 506ms

3.5×

Medium
60 chars · 1.03s

4.2×

Long
156 chars · 2.23s

Inside the 74.8M parameters

56.45M

Synthesizer

8.03M

Speaker Encoder

7.80M

Style Encoder

2.52M

Prosody Predictor

Ships with 5 built-in speakers (NF, SF, NM1, SM, NM2) · Northern & Southern accents · prosody transfer from the reference clip. A lighter Multi-Speaker variant runs at 57.97M params.

Early praise

Runs zero-shot on CPU alone
Fast, real-time-plus generation
Clear Northern / Southern samples
Broad distribution: pip, Docker, Gradio, HF Spaces, Windows .exe

Still to prove

Stability on long texts
Quality consistency across voices
Accent accuracy in the wild
Broad user evaluation still accumulating

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.