Skip to content

orpheus.md

Latest commit

 

History

History
54 lines (39 loc) · 1.46 KB

orpheus.md

File metadata and controls

54 lines (39 loc) · 1.46 KB

Orpheus Engine

OrpheusEngine talks to an OpenAI-compatible completions endpoint and decodes Orpheus audio tokens. In the Zaphod dev-log this was run through LM Studio.

Install

pip install "realtimetts[orpheus]"

The orpheus extra installs snac, but you still need a running compatible model server. Source defaults:

api_url = http://127.0.0.1:1234/v1/completions
model = orpheus-3b-0.1-ft

Minimal Use

from RealtimeTTS import TextToAudioStream, OrpheusEngine, OrpheusVoice


if __name__ == "__main__":
    engine = OrpheusEngine(voice=OrpheusVoice("tara"))
    stream = TextToAudioStream(engine)
    stream.feed("Hello from Orpheus.")
    stream.play()

Source Notes

  • Available source voice names are tara, leah, jess, leo, dan, mia, zac, and zoe.
  • temperature, top_p, max_tokens, and repetition_penalty are sent to the completions endpoint.
  • The prompt format is <|audio|>{voice}: {text}<|eot_id|>.
  • Output is mono 16-bit PCM at 24000 Hz.

Zaphod Dev-Log Notes

  • The working smoke path required LM Studio at 127.0.0.1:1234 with pkmx/orpheus-3b-0.1-ft.
  • Benchmarks showed Orpheus was usable but not among the fastest local paths in that environment.

Troubleshooting

  • If synthesis returns no audio, verify the completions endpoint, model name, and streaming response format.
  • The engine has two synthesize definitions in source; the later definition is the active one in Python.