Palabra × OpenAI Agents SDK

Palabra Realtime STT and TTS as models for the VoicePipeline of the OpenAI Agents SDK. Speech recognition and synthesis go through Palabra instead of gpt-4o-transcribe / tts-1; the agent (LLM, tools, handoffs) stays as it is.

pip install palabra-ai-openai-agents
export PALABRA_API_KEY=plbr_...   # https://platform.palabra.ai/api-keys

Usage

from agents import Agent
from agents.voice import SingleAgentVoiceWorkflow, VoicePipeline
from palabra_openai_agents import PalabraSTTModel, PalabraTTSModel

agent = Agent(name="Assistant", instructions="You are a helpful voice assistant.")

pipeline = VoicePipeline(
    workflow=SingleAgentVoiceWorkflow(agent),
    stt_model=PalabraSTTModel(language="en"),                 # None = server-side language detection
    tts_model=PalabraTTSModel(language="en", voice_id="default_low"),
)

Then use the pipeline as usual: await pipeline.run(AudioInput(buffer=..., frame_rate=...)) for a recorded turn, or StreamedAudioInput for a live microphone. Or hand out both models through the config:

from agents.voice import VoicePipelineConfig
from palabra_openai_agents import PalabraVoiceModelProvider

config = VoicePipelineConfig(model_provider=PalabraVoiceModelProvider(stt_language="en", tts_language="en"))
pipeline = VoicePipeline(workflow=..., config=config)

Examples: examples/voice_agent_mic.py (microphone → agent → speakers, needs sounddevice and OPENAI_API_KEY) and examples/voice_agent_file.py (WAV question in, WAV answer out).

How it maps

Options of PalabraSTTModel: language, input_frame_rate, turn_gap_s, finalize_timeout_s, api_key / client (a configured palabra_ai.Palabra, e.g. for region="us").

Tests

pip install -e ".[dev]"
pytest -m "not live"                          # offline, scripted fake of the Palabra client
PALABRA_API_KEY=plbr_... pytest -m live -s    # against the API; speech is produced by Palabra TTS
PALABRA_TEST_WAV=my.wav PALABRA_TEST_LANGUAGE=ru pytest -m live -k user_recording -s

Notes

• One API Key is not limited to a single live STT session: parallel create_session / transcribe calls each open their own session with the same key.