vox-dia

Dia TTS adapter package for Vox.

Included adapter

• dia-tts-torch — Dia 1.6B text-to-speech backend

Install

pip install vox-dia

Runtime Dependencies

Dia uses an isolated target runtime at $VOX_HOME/runtime/dia for backend packages that must not leak into the base Vox app environment.

The adapter package itself is intentionally lightweight and does not install Torch. The current Dia backend uses Hugging Face Transformers and requires a Vox runtime/image that already provides PyTorch with CUDA. CPU execution is not supported by the official Dia Transformers runtime used by this adapter.

The current Vox adapter is classified for Linux x86_64 CUDA/Torch runtimes. CPU, ONNX, and Spark/ARM NVIDIA paths are not currently production-supported by this adapter. Upstream Dia 1.6B documentation says the model has only been tested on GPUs with PyTorch/CUDA and that CPU support is future work: https://github.com/nari-labs/dia#hardware-and-inference-speed

Plan for at least 12GiB of usable VRAM budget: upstream reports that the full Dia 1.6B model requires around 10GB of VRAM, and Vox budgets the adapter's 10GB model estimate plus the deployment's configured VRAM headroom. A server started with --max-vram 10GiB --vram-headroom 1GiB will reject Dia at load time.

During vox pull, the adapter verifies or installs the Dia-capable Transformers runtime into $VOX_HOME/runtime/dia. It uses the released transformers==4.57.6 runtime rather than a moving source checkout, so clean pulls are reproducible. Model weights remain in the normal Vox model store.

Use with Vox

vox pull dia-tts:1.6b

Dia prompt control is text-driven. Use speaker tags and non-verbal markers in the input text, for example:

[S1] This is Dia speaking. (laughs) [S2] Keep the tags sparse or artifacts can appear.

The adapter supports Dia's audio-prompt voice cloning path when Vox resolves a stored cloned voice into both reference_audio and reference_text. Dia needs the reference transcript for the reference clip; requests with reference audio but no reference text are rejected clearly.

The adapter also exposes Dia generation parameters through Vox synthesis params:

• max_new_tokens (integer, default 3072, range 1..8192)
• guidance_scale (number, default 3.0, range 0..10)
• temperature (number, default 1.8, range 0..3)
• top_p (number, default 0.9, range 0..1)
• top_k (integer, default 45, range 0..200)