Audio8 TTS Preview 0.1B: A 290M-Parameter Zero-Shot Voice Model on Hugging Face

What happened

Audio8 TTS Preview 0.1B landed on the Hugging Face hub on August 23, 2026, under the account suliniacats. It is a zero-shot text-to-speech model with a ~290M parameter footprint (170M main model plus a 120M neural audio codec) that outputs 44.1 kHz audio.

Context

The model card positions this against multilingual TTS systems ranging from 0.6B to 8.5B parameters, with the 0.1B checkpoint aimed at making zero-shot TTS practical at a small footprint. A companion 0.6B model is already published. The repository bundles the v4 mixed checkpoint, the neural codec, tokenizer, processor, and custom Transformers remote code in a single Hugging Face card. The CV3 evaluation table in the model card is drawn from the 0.6B model's published scores and is described as reference comparisons, not a matched re-evaluation of this checkpoint. At collection time the hub page showed zero likes and zero downloads.

How it works

The architecture is Audio8 Falcon H1 with two autoregressive branches. The slow branch (24 layers, width 512, 8 attention heads, 2 KV heads) predicts semantic tokens from text. The fast branch (4 layers, same width and head configuration) predicts codec codebook indices conditioned on the slow branch's hidden state. Acoustic output uses 10 codebooks with 4,096 entries each, decoded by the separate 120M-parameter codec at 44.1 kHz with 2,048 samples per frame, roughly 21.5 frames per second.

For zero-shot voice cloning you supply a reference audio clip and its transcript; omit both and the model uses a default voice. Maximum context is 2,048 packed text and audio positions, so longer scripts need chunking. The model lives in the transformers library with custom remote code, so loading requires trust_remote_code=True. Dependencies: Python 3.11 or newer, torch at least 2.5.0, transformers at least 4.57.0, and a CUDA-capable GPU.

Our read

The CV3 word-error rates are the most telling data. English at 3.307 and Chinese at 3.619 are competitive with systems ten times the parameter count, but Italian at 14.480 and French at 12.349 land in the range where a listener will notice the model struggling. The model card flags the six non-zh/en languages as experimental, and that is the honest read.

A subtler issue: those CV3 numbers come from the 0.6B model's published evaluation table, not a matched re-run on this 0.1B checkpoint. The model card says this explicitly. The quality gap between 0.1B and 0.6B is therefore not quantified anywhere in the sources. The "v4 mixed checkpoint" label is also unexplained; I do not know what training mixture or strategy it refers to.

For a studio, the custom remote code requirement is a practical supply-chain concern. trust_remote_code=True means arbitrary Python executes at load time. A standard safetensors checkpoint does not carry that risk. And at collection time the hub page showed zero likes and zero downloads: there is no community bug tracker, no issue queue, no user-reported edge cases to fall back on. You are the first user in every language except the two primary ones.

What this changes

For a studio running ComfyUI on its own GPU stack, the ~290M total footprint is small enough to load alongside video-generation models without swapping VRAM. The 44.1 kHz codec output is broadcast-quality; no up-resampling pass is needed. Zero-shot cloning from a short reference clip is directly useful for a consistent narrator voice across a series of AI-produced videos.

But there is no ComfyUI-native node. You would write a Python wrapper that calls the transformers pipeline, saves a WAV, and feeds it into the video stack. The 2,048-token context window means scripts longer than roughly 30 seconds of speech will need chunking. If your client brief is in English or Chinese, this is a viable tool to try on Monday. If it is in Italian or French, it is not.

License

The Hugging Face tag reads "license:other" and the model card does not state a licence name. No source names Apache-2.0, MIT, or any other specific licence. Commercial-use rights, redistribution terms, and permissions for derivative models are unconfirmed. Before shipping client or public-facing video content that includes audio generated by this model, check the model card and the GitHub repository for any terms that may not have been reflected in the hub tag.

Key takeaways

• ~290M total parameters (170M main + 120M codec) fits on a single mid-range GPU alongside video models; 44.1 kHz output requires no up-resampling.
• Zero-shot voice cloning is the primary use case; omit the reference clip and transcript for a default voice.
• CV3 error rates are strong for Chinese (3.619) and English (3.307) but high for Italian (14.480) and French (12.349); six languages are explicitly experimental.
• The model requires trust_remote_code=True and custom Transformers code, a supply-chain step beyond a standard safetensors load.
• No licence is named in any source. Commercial use is unconfirmed.

Sources

1. suliniacats/Audio8-TTS-Preview-0.1b — text-to-speech on Hugging Face — tier 3