Overview
Audio8 TTS 0.6B is an open text-to-speech model from Edge0, named exactly that on its product page. The main model has 601M parameters, excluding the neural audio codec shipped alongside it. Give it one reference recording plus its transcript and it reads new text in that voice, keeping the timbre across languages without any per-speaker training.
The architecture is a two-layer DualAR. The slow autoregressive stage emits one semantic token per frame, setting alignment, pauses and the shape of the sentence, while the fast stage reads its hidden states and fills the acoustic codebooks along RVQ depth within that frame. Splitting long-range structure from local detail means the two never compete for one token sequence. The codec runs at 44.1kHz with 2,048 samples per frame, about 21.5 frames per second, so the slow layer makes around 21.5 semantic decisions a second and the fast layer adds ten layers of residual detail to each.
Edge0 lists 11 languages: Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish and Spanish. A Chinese reference can drive an English sentence and the reverse. The repository ships an SGLang Omni adapter with paged attention, dynamic batching, reference audio encoding and an OpenAI-compatible /v1/audio/speech endpoint, and a single H20 generates 128 frames in 0.691 seconds at p50 with an RTF of 0.116 on one stream.
Key Features
- Zero-shot cloning: One reference recording plus its transcript reproduces a voice, with the transcript treated as a condition rather than a note, so clean audio and accurate text give the most stable output.
- Two-layer DualAR: The slow layer handles alignment, pauses and sentence-level shape while the fast layer fills ten RVQ layers within each frame.
- 11 languages: Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish and Spanish, with timbre preserved across languages.
- 44.1kHz codec: Waveforms are rebuilt at 44.1kHz with 2,048 samples per frame, where early codebooks set structure and later ones add timbre and articulation.
- Serving out of the box: An SGLang Omni adapter covers paged attention, dynamic batching and reference encoding, exposing an OpenAI-compatible speech endpoint.
- 601M footprint: A 0.6B main model aimed at edge and consumer hardware, suited to local batch generation or in-product synthesis.
Use Cases
- Audiobooks and podcasts get narrated in batches while keeping one voice across chapters.
- Short video and ad teams audition several narration takes before booking a voice actor.
- One set of course material becomes audio in 11 languages from a single reference voice.
- Call centers and IVR systems standardise on a cloned voice, with script changes handled as text edits.
- Games generate character lines in bulk and add new ones as the story changes.
- Organisations with sensitive recordings synthesise on internal machines so audio never leaves the network.
Pros
- Zero-shot cloning means one clip swaps the voice, with no per-speaker training.
- Two-layer DualAR keeps long sentences stable in rhythm and pauses.
- 11 languages share one voice across cross-lingual output.
- A 44.1kHz codec delivers output at usable resolution without upsampling afterwards.
- The SGLang Omni adapter and OpenAI-compatible endpoint lower integration cost.
- At 601M the main model fits on consumer GPUs and edge devices.
Pricing
According to the release material, weights and code are published under Apache 2.0 for commercial use and further training. For inference, Edge0 documents an SGLang Omni deployment path with paged attention, dynamic batching, reference audio encoding and waveform decoding behind an OpenAI-compatible /v1/audio/speech endpoint; a single NVIDIA H20 generates 128 frames in 0.691 seconds at p50 with an RTF of 0.116 on one stream, excluding cold start and compilation. No hosted API or per-character pricing has been published. On Seed-TTS the model reports 1.506 English WER with 63.2 similarity and 0.950 Chinese CER with 73.1 similarity. Obtain the speaker's consent under local law before cloning a voice.
Summary
Voice cloning has always been gated by two steps: custom timbre requires training, and training requires compute. Audio8 compresses both into a single forward pass where one recording is the entire setup. Its trade-off is legible too, using two autoregressive layers to separate sentence planning from in-frame detail and a 44.1kHz codec so the output needs no second pass, which leaves the real limit at how clean your reference clip is and whether its consent is in order.
Version History
- Audio8 TTS 0.6B listed: two-layer DualAR zero-shot cloning (2026-10-31): Form at listing: Edge0's open text-to-speech model with a 601M-parameter main model excluding the bundled neural audio codec. Two-layer DualAR separates sentence-level planning from in-frame RVQ detail, with a 44.1kHz codec at 2,048 samples per frame and about 21.5 frames per second. One reference recording plus its transcript enables zero-shot cloning across 11 languages: Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish and Spanish. The repository ships an SGLang Omni adapter and an OpenAI-compatible speech endpoint, generating 128 frames in 0.691 seconds at p50 on a single H20 with an RTF of 0.116.