Parameter-count clarification: is Audio8 0.6B or approximately 0.94B end-to-end?
I checked the released weights because the model card repeatedly describes Audio8 TTS as a 0.6B model.
The 0.6B figure is accurate for model.safetensors: Hugging Face reports 601,159,424 parameters. However, the released inference pipeline also requires the bundled codec.pth. Instantiating the codec from the project’s published code and configuration gives another 337,428,866 learned parameters.
That makes the complete shipped text-to-waveform stack approximately:
601,159,424 + 337,428,866 = 938,588,290 parameters
The model card does say “main model ... excluding the codec” in its architecture table, so the individual number is disclosed. However, the headline and benchmark tables still label the overall TTS release as 0.6B, while describing the repository as a complete checkpoint with its neural codec and zero-shot voice cloning.
Could the team clarify whether “0.6B” is intended to mean only the DualAR backbone? It would be helpful to label benchmark tables as main-model parameters, codec excluded, and provide the approximately 0.94B end-to-end count alongside it.
The model may still be technically impressive. This is specifically about making the footprint comparison unambiguous.
Thanks for the careful analysis and for raising this point.
You’re correct that the 0.6B figure refers specifically to the DualAR backbone, following the naming convention commonly used in the industry, where the primary generative model is counted separately from auxiliary components such as neural codecs.
For inference, it’s also worth noting that the full codec is not always required. If the reference audio tokens are prepared in advance (which is the typical deployment scenario for zero-shot voice cloning), only the codec decoder (~120M parameters) is needed during waveform generation. The codec encoder does not participate in the online inference path in this case.
We agree that the current model card could make this distinction clearer. We’ll update the documentation to explicitly distinguish:
- the 0.6B main model (DualAR backbone),
- the codec components,
- the end-to-end deployment footprint under different inference settings.
We’re also working on a more memory-efficient inference pipeline and will release improved inference code and models that better reflect practical deployment scenarios.
Thanks again for the thoughtful feedback—we appreciate it.