Audio8 TTS Preview 0.1B (ONNX INT8)

A compact multilingual text-to-speech model with zero-shot voice-cloning, packaged by audio8 for CPU inference under ONNX Runtime. The routed source is the quantized deployment repo; the FP-precision checkpoint it was built from lives at Audio8/Audio8-TTS-Preview-0.1b and is read here alongside it, because on three separate facts the two cards do not agree.

What it is

DualAR, credited to fish-audio-s2-pro. A slow autoregressive branch predicts one semantic token per audio frame; a fast branch predicts that frame's codec codebooks, conditioned on the slow hidden state and the codebooks already emitted. The slow branch is a Falcon-H1 hybrid — attention plus Mamba (24 layers, width 512, 8 attention heads, 2 KV heads, Mamba d_state 64); the fast branch is 4 layers at the same width. Acoustic tokens are 10 codebooks of 4,096 entries over a 44.1 kHz neural-audio-codec at 2,048 samples per frame (~21.5 frames/s), with a 2,048-position packed text+audio context.

This is the first state-space component in the spoke's TTS corner. Every other architecture here is transformer, flow-matching (luxtts) or diffusion; the Mamba half of the slow branch is what makes the per-token recurrent ONNX export below tractable, since the state is carried explicitly rather than re-read from a growing KV cache.

The quantization split. Slow and fast AR ship as INT8 with dynamic activation quantization (MatMulInteger); the codec encoder and decoder stay FP16; output is FP32 at 44.1 kHz mono. The slow AR is exported as a per-token recurrent graph — attention and Mamba states passed in and out explicitly — so no session holds a growing cache. Online files are ~0.6 GB, the full repo including the voice-registration encoder ~1.0 GB. Apache-2.0 on this repo, and the runtime around it (CLI, web UI, HTTP API, streaming PCM, an OpenAI-compatible /v1/audio/speech endpoint) lives in Audio8-AI/Audio8_TTS — 895 stars, Python, first pushed 2026-07-29.

Speed: fast enough on a CPU, by its own numbers

The card reports ~19 ms per slow-AR token, ~8 ms per fast-AR frame and a one-time prefill of ~19 ms per prompt token, on an unnamed 8-thread host. One audio frame costs one of each, so 27 ms of compute buys 46.5 ms of audio: a real-time factor near 0.58, roughly 1.7× faster than real time on CPU — our arithmetic from the card's two numbers, not a figure the publisher printed. Memory is given as about 0.6 GB after the online sessions load.

That is the interesting claim. supertonic-2 gets a far better RTF (0.012 on an M4 Pro CPU) but cannot clone a voice at all; kokoro is smaller and also cannot. Audio8 is the first model in this wiki to put zero-shot cloning at 44.1 kHz above real time on a CPU with no GPU, which is the combination open-weight-tts's size ladder had been trading away at the small end.

The benchmarks, and the axis they lose on

From the base card. Lower WER/CER is better, higher SIM better; the publisher notes different normalizers make these reference comparisons rather than a matched ranking, and that fish-audio-s2-pro and Higgs Audio v2 were re-evaluated locally because published values were unusable or absent.

Seed-TTS:

Read the two columns separately, because they disagree. On intelligibility the 0.17B model is within 0.06 WER of a 4.6B one — near-parity at about a twenty-seventh of the parameters. On speaker similarity it is last but one: 56.7 against Fish's 64.6 and IndexTTS2.5's 82.3, a gap of 8 to 26 points. Since the product being sold is voice cloning, the column the model loses is the column that matters, and a headline that quotes only WER would invert the result. This is the tts-benchmarks problem in one table.

The CV3 error-rate table adds the language shape: zh 3.619 and en 3.307 are competitive with models ten times the size, while it 14.480, fr 12.349 and ja 12.322 are two to three times the 0.6B sibling's. The compact checkpoint holds up in its two primary languages and degrades sharply outside them.

Three places the two cards contradict each other

Same publisher, same model, two repositories, published six days apart.

The parameter gap is the one to hold onto: neither number is 0.1B, and the name is below both. The language gap runs the other way from what the base card permits — it says in as many words that "evaluation coverage does not expand the Preview checkpoint's supported-language claim," and then the derived repo lists three languages the base card never mentions, none of them evaluated in either table. Memory is plausibly just two different hosts. The licence line is the substantive one, and it gets its own section.

Recorded, not resolved. No attempt here to decide which side is right; both are quoted with their source.

A third licence shape: revenue-capped

open-weight-tts has carried a two-way split since June — permissive (Apache/MIT) versus research-only or paid-commercial (fish-audio-s2-pro). The Audio8 Community License v1.0 on the base checkpoint is neither:

• Non-commercial use (research, personal, educational, evaluation) — free.
• Commercial use — free, provided the entity's annual revenue is under US$2,000,000, counted consolidated across parent companies, subsidiaries and affiliates.
• At or above US$2M, commercial use is not permitted under the licence; a separate written licence from Audio8 is required.

The licence also carries behavioural terms: no impersonation without consent, no deceptive content, and disclosure of synthetic content where required. That matters for the absence synthesis tracks — the census in free-voice-clone-list found no consent, watermark or misuse language anywhere across 53 models, and the one licence there with behavioural restrictions (OpenRAIL-M on supertonic-2) belonged to the one model that cannot clone. Audio8 is the second data point, and unlike Supertonic it does clone. Both model cards also repeat the consent and disclosure language in prose.

The gap the two repos leave. The ONNX repo states Apache-2.0 over "the code and model weights," while the weights it quantized come from a repo under the revenue-capped licence. The cards do not address the relationship. Anyone above the revenue line who reads only the ONNX card would conclude they have a permissive licence; the base card says otherwise. Flagged as an observed discrepancy in the published terms — this wiki is not reading it as a legal conclusion.

Quantization, unmeasured

The card is unusually direct that it did not measure the cost: INT8 "can change sampled token sequences, so quality should be evaluated for each target language, voice, and deployment setting rather than assuming bit-for-bit equivalence," and FP32 per-token graphs ship alongside for faithful decoding. No WER, CER or SIM delta between INT8 and FP32 is published anywhere in either repo.

That sits directly against the spoke's other measurement of the same question: on-device-streaming-asr-study found INT8 measuring marginally better than FP32 on ASR, with the accuracy cost living in the PyTorch→ONNX export rather than in the precision. Two things worth keeping apart — that study measured a discriminative task with a single correct answer, while sampling from a generative token sequence is where a small logit perturbation can compound. So the publisher's caution is reasonable, and it is also unfalsifiable as shipped. Cross-wiki: quantization.

Provenance of the port

The ONNX design — stacked recurrent interface, valid-prefix attention, and the INT8 recipe — is credited as adapted from a community conversion, Masterx/Audio8-TTS-Preview-0.1B-ONNX-INT8, with the graphs re-produced by Audio8 from the official checkpoint rather than reused. A vendor adopting a community port of its own model is the inverse of the luxtts pattern, where a third party distilled someone else's model and the derivative outgrew the base.

Reception and caveats

17 likes and 0 downloads on the ONNX repo the day it appeared; the base checkpoint has 3,640 downloads and 154 likes since 2026-08-19. No independent evaluation exists yet — every number on this page is the publisher's own, and the Elo boards this wiki tracks (tts-arena-leaderboard) do not list it. The publisher's own stated limits: preview status, limited dialect coverage, and degraded stability from long, noisy or mistranscribed reference audio.

Re-checked 2026-08-30 (absence-claim drain): still nothing independent. Searching turns up only Audio8's own repos, Space, GitHub and listening-preview page, plus a 0.6B evaluation table the 0.1B card borrows as a "reference comparison rather than a strictly matched re-evaluation" — the publisher comparing its small model against its own bigger one. The claim holds.

Related

audio8 · text-to-speech · open-weight-tts · voice-cloning · fish-audio-s2-pro · neural-audio-codec · tts-benchmarks · kokoro · supertonic-2 · kittentts · luxtts · on-device-streaming-asr-study · quantization