Audio8TS: a tiny 1.7GB text-to-speech model for local voice cloning

Audio8TS is an ultra-compact text-to-speech engine that packs zero-shot voice cloning into a tiny 1.7GB package, designed to run entirely offline on standard consumer PCs.

While modern voice cloning setups typically demand massive multibilion-parameter checkpoints and high-end GPUs, Audio8TS takes the opposite path. Its main generative model sits at roughly 170 million parameters, and the dedicated audio codec decoder adds around 120 million parameters, keeping the complete footprint around the 0.1B mark.

What it can do

The primary use case is zero-shot voice cloning. Feed it just a few seconds of clean reference speech alongside the corresponding transcript, and it immediately captures the speaker's tone to voice any text you throw at it.

English and Chinese are the primary supported languages, delivering solid pronunciation and strong timbre matching. It also includes experimental multilingual evaluation support for German, Spanish, French, Italian, Japanese, and Korean.

Despite the tiny footprint, evaluation benchmarks indicate decent character error rates and speaker similarity scores. It won't beat massive studio models on every nuance, but for a 1.7GB bundle, the output is surprisingly coherent.

How to run it

Under the hood, Audio8TS runs an Audio8 Falcon H1 architecture split into slow and fast autoregressive branches. The slow branch predicts semantic tokens, while the fast branch predicts codec codebooks conditioned on the slow hidden state before the neural codec translates tokens into raw waveforms.

Accessibility is the real highlight here. Aside from standard CUDA GPU inference via Hugging Face Transformers, the authors packaged an ONNX INT8 build tailored specifically for resource-constrained CPUs.

In the CPU configuration, the slow and fast autoregressive engines operate on per-token INT8 graphs, while the neural codec uses FP16. Voice registration sessions release memory before synthesis begins, preventing memory spikes on basic laptops.

The complete model weights, codec, tokenizer, processor, and inference scripts are open on Hugging Face under Audio8/Audio8-TTS-Preview-0.1b. It requires Python 3.11 or newer along with PyTorch and Transformers.

Why it matters

If you want to bake custom voice generation into local desktop apps or edge hardware without paying cloud API bills, Audio8TS is definitely worth a test run. The hardware barrier is practically zero.

We think Audio8TS takes a very smart compromise. Instead of chasing state-of-the-art benchmarks with billions of parameters, it proves that disciplined architecture choices can deliver usable voice cloning at the 100M scale. It sets a clean, realistic benchmark for local-first audio AI.