isolate model inference configs

[pgrayy]

Description

Isolating the model inference configs to more easily extract them from provider_configs.

Testing

• I ran hatch run bidi:prepare: Updated unit tests
• I ran the following test script:

import asyncio
import json

from strands.experimental.bidi import BidiAgent
from strands.experimental.bidi.io import BidiAudioIO, BidiTextIO
from strands.experimental.bidi.models import BidiNovaSonicModel
from strands.experimental.bidi.models.gemini_live import BidiGeminiLiveModel
from strands.experimental.bidi.models.openai_realtime import BidiOpenAIRealtimeModel
from strands.experimental.bidi.tools import stop_conversation

from strands_tools import calculator


async def main() -> None:
    model = BidiNovaSonicModel(
        model_id="amazon.nova-sonic-v1:0",
        provider_config={
            "audio": {
                "voice": "matthew",
            },
            "inference": {
                "max_tokens": 600,
            },
        },
        client_config={"region": "us-east-1"},
    )
    # model = BidiGeminiLiveModel(
    #     model_id="gemini-2.5-flash-native-audio-preview-09-2025",
    #     provider_config={
    #         "audio": {
    #             "voice": "Charon",
    #         },
    #         "inference": {
    #             "max_output_tokens": 600,
    #         },
    #     },
    #     client_config={"api_key": "..."},
    # )
    # model = BidiOpenAIRealtimeModel(
    #     model_id="gpt-realtime",
    #     provider_config={
    #         "audio": {
    #             "voice": "coral",
    #         },
    #         "inference": {
    #             "max_output_tokens": 700

    #         },
    #     },
    #     client_config={"api_key": "..."},
    # )
    agent = BidiAgent(model=model, tools=[calculator, stop_conversation])

    audio_io = BidiAudioIO()
    text_io = BidiTextIO()
    await agent.run(inputs=[audio_io.input()], outputs=[audio_io.output(), text_io.output()])

    print(f"MAIN - stopping agent: {json.dumps(agent.messages, indent=2)}")


if __name__ == "__main__":
    asyncio.run(main())

• Model responded as expected.
• Toggling the max tokens did affect out size.
• Stop conversation tool worked as expected.
• Interruption worked.

[pgrayy]

Voice is passed in through the "audio" config.

[pgrayy]

No need to explicitly provide defaults. Nova already has implicit defaults for these that we can rely on.

[pgrayy]

Voice provided through "audio" config of type AudioConfig.

[pgrayy]

• type always has to be realtime and is already set by us.
• instructions is set by us through system prompt.
• voice is set by us through "audio" config.
• tools is set by us through the passed in tools param.
• input_audio_format, output_audio_format, input_audio_transcription, and turn_detection are not top-level configs and so would lead to exceptions if setting.

For more details on supported settings, see https://platform.openai.com/docs/api-reference/realtime-client-events/session/update#realtime_client_events-session-update-session.

[pgrayy]

Using to promote consistency. We use snake_case everywhere else.

[mehtarac]

i think we should have a default voice here

[pgrayy]

It does work without specifying. I tested that on all the models actually. With that said, we could remove the default voice setting on all configs but I didn't want to make too many changes here.