OpenAI Agents SDK (voice) - Lokutor

The Agents SDK’s built-in OpenAI voice models cannot be redirected to Lokutor with a base URL alone: the TTS model requests raw 24 kHz PCM (Lokutor always returns a 44.1 kHz WAV) and transcribe() uses the Realtime WebSocket API. The lokutor-agents-voice adapter calls /v1/audio/speech and /v1/audio/transcriptions and converts the audio.

pip install git+https://github.com/lokutor-ai/lokutor-agents-voice
export LOKUTOR_API_KEY=...

from agents.voice import VoicePipeline, SingleAgentVoiceWorkflow
from lokutor_agents_voice import LokutorVoiceModelProvider

provider = LokutorVoiceModelProvider(voice="F1", language="en")
pipeline = VoicePipeline(
    workflow=SingleAgentVoiceWorkflow(agent),
    stt_model=provider.get_stt_model(None),
    tts_model=provider.get_tts_model(None),
)

Limits

Push-to-talk input only (AudioInput). StreamedAudioInput is not implemented; use /ws/stt directly.
One synthesis request per text chunk; audio is resampled to 24 kHz with linear interpolation.

Verification

Run against the live API with openai-agents 0.24.0: a TTS to STT round trip and a VoicePipeline with an echo workflow. Not run: SingleAgentVoiceWorkflow with an LLM, and streamed input.