vox-orpheus 0.1.7

Orpheus TTS adapter for Vox

vox-orpheus

vox-orpheus provides a Vox TTS adapter for Orpheus.

Adapters:

• orpheus-tts-vllm - Orpheus medium 3B backend through vLLM

Install

pip install vox-orpheus

Runtime Dependencies

The adapter package is intentionally light. The orpheus-speech backend and its vLLM/SNAC runtime dependencies are installed on demand into the isolated target runtime $VOX_HOME/runtime/orpheus.

During vox pull, the adapter verifies or installs the Orpheus runtime without loading model weights. Model weights remain in the normal Vox model store.

Use with Vox

vox pull orpheus-tts:medium-3b
vox run orpheus-tts:medium-3b "Hello from Orpheus" --voice tara
vox run orpheus-tts:medium-3b "<happy>Hello from Orpheus.<laugh>" --voice tara

Preset voices:

• tara
• leah
• jess
• leo
• dan
• mia
• zoe
• zac

Orpheus supports expressive prompt markup in the text itself. Use tags sparingly; dense markup can create artifacts or unstable delivery.

Non-verbal tags:

• <laugh>
• <chuckle>
• <sigh>
• <cough>
• <sniffle>
• <groan>
• <yawn>
• <gasp>

Phrase-level emotion tags:

• <happy>
• <sad>
• <angry>
• <excited>
• <fearful>
• <surprised>
• <disgusted>
• <neutral>

This adapter supports Orpheus preset voices, not reference-audio voice cloning. Requests with reference_audio or reference_text are rejected clearly rather than silently ignored.

The adapter exposes Orpheus generation parameters through Vox synthesis params:

• temperature (number, default 0.6, range 0..2)
• top_p (number, default 0.8, range 0..1)
• repetition_penalty (number, default 1.1, range 1..2)
• max_tokens (integer, default 1200, range 1..4096)

The current Orpheus backend is vLLM-based and is classified for Linux x86_64 CUDA runtimes. CPU and Spark/ARM NVIDIA execution are not currently supported by this adapter because the orpheus-speech/vLLM runtime is not packaged as a portable CPU or ARM NVIDIA backend. Plan for at least 10GiB of usable VRAM budget before deployment headroom.