English · Preview 0.6B
The earlier release, using the same English passage as the opening player.
Audio8 TTS.
Hear the detail.
Compact speech models for expressive narration and zero-shot voice cloning, with official samples you can hear before exploring the code.
Try It NowIndependent model guide. Not affiliated with the Audio8 team.
Realistic 0.6BRecorded sample · English
“Listen up! You think this is a maybe? You need to be on that screen, and you need it now!”
Try Audio8 TTS
Explore text-to-speech and voice cloning in the official Hugging Face playground.
Voicv
Try Voicv ↗
Make your own voiceover with an online AI voice studio.
This embedded demo is hosted by Edge0 on Hugging Face. Text and audio you submit are processed by that service. Availability depends on its queue and GPU allowance; sign-in may be required.
Listen, then compare
Start with the same passage in two releases. Listen for pauses, emphasis, and the way a sentence finishes.
Same text.
Different delivery.
These players use recordings published by the Audio8 project. They do not generate new audio or upload your voice. The examples illustrate selected outputs, rather than a guarantee that every script will sound the same.
Use headphones, keep playback volume consistent, and compare one sentence at a time. Pay attention to whether the delivery serves the meaning, rather than simply sounding dramatic.
Explore the full official collectionEnglish · Realistic 0.6B
Compare its emphasis and rhythm with Preview using the same source passage.
Reference voice
The project’s original reference recording provides the voice context for both outputs.
Chinese · Realistic 0.6B
Hear a Chinese example from the official collection, then compare phrasing across languages.
Choose your model
The name matters: Realistic and Preview have different language coverage and deployment options.
Realistic 0.6B
The current 0.6B release focuses on more natural delivery and supports 23 languages. Its parameter count excludes the separate audio codec. Start here when you want to explore the newer voice quality.
The official setup uses an explicit Realistic model identifier. There is currently no official Realistic ONNX export, so a Preview deployment recipe should not be assumed to apply.
Read the Realistic model cardPreview 0.6B
The earlier 0.6B checkpoint supports 11 recommended languages. It remains useful for reproducing earlier examples and comparing changes in speech delivery. Its ONNX INT4 deployment route is documented separately.
Keep the model version alongside saved settings and test scripts. Otherwise, a change of checkpoint can be mistaken for an improvement caused by a prompt or recording adjustment.
Browse the release documentationPreview 0.1B
This smaller model family targets a compact deployment footprint, with Chinese and English as its primary languages. The dedicated ONNX INT8 runtime offers CPU inference, streaming, and local serving.
Choose it for a hardware experiment only after checking the exact runtime and model license. A smaller model name does not establish a universal speed or memory requirement.
Explore the compact model23 ways to be heard
Realistic expands the earlier language set. Coverage describes supported languages, not equal quality across accents, subjects, and recording conditions.
Arabic, Cantonese, Chinese, Dutch, English, Finnish, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Thai, Turkish, and Vietnamese.
Test a real script
Choose an evaluation that resembles the work you actually want to produce.
Narration and lessons
For an explainer or lesson, test a passage with terminology and a change of topic. A pleasant opening is not enough: check whether emphasis helps a listener follow the argument and whether punctuation creates useful pauses. Keep a written list of pronunciation corrections for later takes.
Dialogue and localization
For dialogue, compare the intention of a line with its delivery before assembling several turns. For localization, ask a fluent speaker to review wording and pronunciation in each target language. Translating a script and synthesizing a voice are separate steps; a convincing sound does not establish that the translation is accurate.
Prepare a better take
A short, repeatable test makes it easier to hear what changed before committing to a longer recording.
Write for listening
Read your script aloud before synthesis. Replace dense sentences with clear phrases, spell out ambiguous numbers, and decide how names should sound. Begin with a short paragraph that includes the difficult parts of your actual project.
Choose a clean reference
For voice cloning, use a recording you have permission to use and provide its accurate transcript. Prefer one speaker without overlapping voices or background music. Check that the transcript matches what is spoken, including short words and pauses.
Select the release
Follow the official installation guide for your chosen checkpoint. The Realistic quick start recommends Python 3.10 or newer and a CUDA-capable GPU. Select the model explicitly instead of relying on a script’s default, which may still point to Preview.
Review the whole result
Check pronunciation, missing words, breath placement, and changes in loudness. Listen through the end of the output. Save your script, model name, reference, and settings together so a useful result can be reproduced and compared later.
Before you begin
What is Audio8 TTS?
Audio8 TTS is a family of compact text-to-speech models for speech generation and zero-shot voice cloning. This independent guide collects listening examples and links to the project’s release documentation. Model weights and inference tools are maintained by the upstream project.
Can I generate audio here?
Yes, through the embedded official Hugging Face demo. The upstream Space handles generation and may require sign-in or have usage limits. This site does not operate an inference server. The separate listening players contain prerecorded examples; playing them does not submit a generation request.
What does zero-shot mean?
In this context, a model can condition speech on a reference voice without a separate training run for that speaker. Reference quality and transcript accuracy still matter. Similarity can vary, so listen carefully rather than assuming a perfect match.
Which release runs on a CPU?
The project documents a dedicated CPU ONNX INT8 runtime for Preview 0.1B and an ONNX route for Preview 0.6B. Realistic has different deployment requirements. Check the guide for the exact release, hardware, and format you intend to use.
Are the samples a ranking?
No. The upstream listening collection contains selected examples. It is useful for noticing delivery differences, but it does not replace a matched evaluation on your own scripts. Language support, intelligibility, speaker similarity, and perceived naturalness answer different questions.
What should I test first?
Try a representative passage containing a name, a number, and a longer sentence. Compare repeated outputs at a consistent volume. Review the exact model’s license and the permissions for your reference recording before publishing generated speech.