Open-source model, on-device

Audio8 TTS

Run the Audio8-TTS-Preview-0.6b model right here in your browser. Type text, press generate, and the 0.6B DualAR model speaks it locally in 44.1 kHz audio. No signup, no server, no upload.

Your text

0 / 500
Model not loaded on this device yet.

Speech model

Audio8 TTS Preview Runs on your device

Open-source 0.6B DualAR model with 44.1 kHz output. Generation happens locally; your text never leaves this page.

Audio8 follows the written text; it has no hidden language-ID input. This selection sets writing guidance and accessibility metadata.

Profiles stay in this browser. A clone usually gives more consistent output than the unconditioned preview.

0.7

Lower values sound steadier, higher ones add variation. 0.7 matches the official default.

Model weights: Audio8/Audio8-TTS-Preview-0.6B-ONNX-INT4, Apache 2.0. Weights stream from the Hugging Face CDN straight into your browser cache.
The quick answer

Audio8 TTS is an open-source 0.6B speech model with 11 recommended languages and zero-shot voice cloning. This page caches about 572 MiB of synthesis files, tries native WebGPU before a threaded CPU fallback, and keeps text, reference voices, profiles, and generated 44.1 kHz audio on your device.

The model

What is Audio8 TTS?

Audio8 TTS Preview 0.6B is a compact text-to-speech model published by the Audio8 team on Hugging Face. It uses a DualAR design borrowed from Fish Audio S2 Pro: a slow autoregressive transformer plans one semantic token per audio frame, and a smaller fast transformer fills in the ten codebooks a neural codec needs to turn those tokens back into sound.

The result speaks with prosody that holds up across languages, and it does two things most compact models cannot. It outputs 44.1 kHz audio instead of the usual 24 kHz, and it supports zero-shot voice cloning, where a short reference clip plus its exact transcript steers who the output sounds like.

This release is a preview. The team recommends sticking to eleven core languages and says broader coverage and Chinese dialect support are coming. Everything ships under Apache 2.0, so anyone can inspect, fine-tune, or deploy it.

Model facts

Model
Audio8-TTS-Preview-0.6b (open weights)
Parameters
601 million, excluding the audio codec
Architecture
DualAR transformer: a 24-layer slow model and a 4-layer fast model
Audio codec
Neural codec, 10 codebooks, 44.1 kHz mono output
Frame rate
About 21.5 frames per second, 2,048 samples per frame
Context window
Up to 2,048 packed text and audio positions
Languages
11: English, Spanish, French, German, Italian, Dutch, Polish, Japanese, Korean, Chinese, Cantonese
Voice cloning
Zero-shot cloning from a reference clip plus its transcript
License
Apache 2.0 for code and weights
Browser download
About 572 MiB for synthesis; voice cloning optionally adds about 396 MiB
Four steps

How to use Audio8 TTS in your browser

  1. 01

    Open this page on a desktop

    The tool loads straight in the browser. No account, no install, no API key. Phones and tablets need to skip this one, the model wants PC or Mac hardware.

  2. 02

    Let the weights download once

    The first Generate click pulls about 572 MiB of INT4 model files from the Hugging Face CDN. Your browser caches them, so later visits skip the network download.

  3. 03

    Choose a language and voice

    Select one of the 11 recommended languages. Audio8 follows the writing itself rather than a language ID. Use the unconditioned preview or register a permitted 0.5–30 second reference voice locally.

  4. 04

    Listen and download

    Wait for the smooth final result, or enable the optional live preview for provisional chunks. Save a lossless 44.1 kHz WAV or convert it to MP3 inside the browser.

Zero-shot cloning

How Audio8 voice cloning works

Give the model a clean reference recording between half a second and thirty seconds long, plus the exact transcript of what is said in it, and it produces new speech in that voice. No training run, no fine-tuning, no voice marketplace. The bundled codec encodes the reference into the same token space the model generates in, which is why the transcript has to match the recording word for word.

Open Create a local voice clone above to run that workflow in-browser. The optional encoder adds about 396 MiB, then saves only the resulting codec profile in this browser. Registration temporarily releases the synthesis sessions to keep peak memory under control.

Clone responsibly

The upstream model card says it plainly: get consent before cloning a voice, and disclose synthetic audio where listeners would reasonably expect a real person. A cloned voice used without permission is the fastest way this technology gets locked down for everyone.

Honest comparison

Audio8 TTS vs Kokoro

We run Kokoro across the rest of this site, so here is the honest internal comparison: where each model wins, and why both earn a spot.

Where it runs

FreeTextoSpeech

Native WebGPU on compatible Chrome/Edge devices, with threaded CPU/WebAssembly fallback.

Kokoro (our main tool)

In-browser engine too, roughly 300 MB lighter to download.

Output quality

FreeTextoSpeech

44.1 kHz neural codec output with expressive prosody.

Kokoro (our main tool)

24 kHz output, very natural narration for its size.

Language coverage

FreeTextoSpeech

11 languages including Korean, German, Dutch, Polish, and Cantonese.

Kokoro (our main tool)

9 languages focused on English, European languages, Hindi, Japanese, Mandarin, and Portuguese.

Speed on a typical laptop

FreeTextoSpeech

Heavier: an AR model generating about 21.5 frames per second of audio.

Kokoro (our main tool)

Much faster, close to real time on most machines.

First-use download

FreeTextoSpeech

About 572 MiB, plus an optional 396 MiB voice encoder.

Kokoro (our main tool)

A few hundred MB at full precision.

Maturity

FreeTextoSpeech

Preview release with limited dialect coverage.

Kokoro (our main tool)

Stable v1.0 model we serve across the whole site.

Both models are Apache 2.0 open source. Numbers describe the current public releases: Audio8-TTS-Preview-0.6b and Kokoro-82M-v1.0.

Best practices

Getting good output from the Audio8 preview

Small habits separate a demo-grade clip from something you would publish. These six matter most with an autoregressive model.

  • 01

    Keep requests short while you learn the model

    The upstream runtime recommends roughly 150 characters for best quality. Generate one sentence at a time, then stitch the WAVs for longer scripts.

  • 02

    Write punctuation the way you want it spoken

    Commas buy a short beat, periods a longer one, ellipses a held pause. The DualAR model reads these as delivery cues, so a line that lands flat often just needs a comma.

  • 03

    Drop creativity if samples drift

    The creativity slider sets sampling temperature. At 0.7 you get the official default balance. If a long word warbles or a syllable repeats, pull it toward 0.5 and generate again.

  • 04

    Match the language, spell out the edge cases

    Acronyms read better with periods between letters, N.A.S.A. style. Unusual names read better spelled phonetically. Test the tricky word alone, fix the spelling, then run the full passage.

  • 05

    Chrome or Edge can use the GPU

    Current Chrome and Edge are tested for native WebGPU with FP16 support. If the three model probes do not pass, the tool automatically rebuilds with the threaded WebAssembly CPU backend.

  • 06

    Grab the WAV, convert later

    The WAV keeps the full 44.1 kHz signal for editors and DAWs. Only reach for the MP3 button when file size matters more than headroom.

Straight talk

Limitations you should know before relying on it

It is a preview checkpoint

Language coverage is deliberately narrow in this release, and quality varies per language. The upstream card recommends evaluating output for your target language rather than assuming parity with larger deployments.

Hardware decides your wait

Compatible Chrome and Edge devices try native WebGPU first. Other systems use threaded CPU/WebAssembly where isolation is available. Either path can take time because every audio frame requires eleven autoregressive model calls.

Long scripts need sectioning

The packed context spans 2,048 positions shared between prompt and generated audio. Work a paragraph at a time and join the WAVs in an editor for long-form narration.

Synthetic audio carries responsibility

The upstream maintainers flag impersonation and misinformation risks explicitly. Use consented voices, label synthetic narration where disclosure matters, and follow the laws that apply to you.

FAQ

Audio8 Tts FAQs

01

What is Audio8 TTS?

Audio8 TTS is an open-source text-to-speech model family from the Audio8 team. The current preview, Audio8-TTS-Preview-0.6b, packs 601 million parameters into a DualAR architecture inspired by Fish Audio S2 Pro. It speaks 11 languages and supports zero-shot voice cloning, and both code and weights ship under Apache 2.0.
02

Is Audio8-TTS-Preview-0.6b free to use?

Yes. The model is released under Apache 2.0, which allows commercial use without royalties. This page runs it on your own device, so there is no usage quota from us either. Check the upstream NOTICE file in the official repository for attribution details.
03

Does my text get uploaded anywhere?

No. Generation and optional voice registration happen inside your browser using WebGPU or WebAssembly. The Audio8 engine fetches its runtime and model files from the Hugging Face and jsDelivr CDNs. Text, reference audio, saved voice profiles, and generated audio never touch our servers; the site may still send ordinary usage telemetry without that content.
04

Why is the first generation slow?

Two reasons. First, the browser downloads about 572 MiB of synthesis files once. Second, this autoregressive model performs one Slow AR call and ten Fast AR calls for every audio frame. Compatible Chrome and Edge devices try native WebGPU; other systems use the threaded CPU fallback.
05

Can it clone a voice in the browser?

Yes. Open Create a local voice clone, provide a permitted 0.5–30 second recording and its exact transcript, and download the optional 396 MiB encoder. The resulting codec profile stays in IndexedDB on that browser; the original recording is not stored by the tool.
06

Which languages does Audio8 TTS support?

The preview targets eleven languages: English, Spanish, French, German, Italian, Dutch, Polish, Japanese, Korean, Chinese, and Cantonese. The team states that broader multilingual coverage and Chinese dialect support are planned for future releases.
07

Audio8 or Kokoro, which should I use?

Use Kokoro, our main tool, when you want fast generation, 54 ready voices, phone support, and proven stability. Pick Audio8 when you want to try the newest open model, need 44.1 kHz output, or work in German, Korean, Dutch, Polish, or Cantonese. Many people keep both bookmarked.
08

Does it work offline?

Model weights and voice profiles are cached locally, so repeat generations do not re-download them. A complete offline reload still depends on your browser retaining this page and its runtime modules; clearing site data removes the cache and local voices.

Still wondering? Get in touch →

Keep going

Related tools

Main generator ↗

Free Text to Speech

54 Kokoro voices in 9 languages, instant cloud generation, WAV download.

Voice quality ↗

Natural AI Voice Generator

Why modern neural voices sound human and how to pick one.

Format ↗

Text to MP3

Convert text to ready-to-share MP3 at 128 to 320 kbps.

On-device option ↗

Text Reader Online

Read pasted text aloud with the voices already on your device.

Try it now

Try the newest open voice model.

Runs in your browser, free, private.

Open the tool See all voices
visual guide showing text input, AI voice selection, waveform generation, and audio download

Visual guide

A visual guide to turning text into downloadable speech audio.