feat: kokoro tts support

[MagdalenaKotynia]

Purpose

• To support the usage of Kokoro-TTS model. Kokoro-TTS was selected based on its high-quality speech output, small size, and potential to run on edge devices (it is in ONNX format)

Proposed Changes

• Developed a class implementing the TTSModel interface for the Kokoro-TTS model.
• Updated docs with newly supported model.
• Updated example with TTSAgent to be able to use the newly supported model

Testing

• Source ros2
• Install RAI following the instructions for developers from docs
• Install s2s package: poetry install --with s2s

With TTSAgent

• Run TTSAgent example: python examples/s2s/tts.py
• In another terminal run the following script to send ros2hri message to ros2 topic:

from rai.communication.ros2.connectors import ROS2HRIConnector
from rai.communication.ros2.messages import ROS2HRIMessage
import rclpy
import time

rclpy.init()
my_hri_msg = ROS2HRIMessage(
    text="Hello, human! This is a test message. How are you?",
    message_author="ai",
)

hri_connector = ROS2HRIConnector()

hri_connector.send_message(
    message=my_hri_msg,
    target="/to_human"
)

try:
    print("Sending message... Press Ctrl+C to exit")
    time.sleep(10)
    
except KeyboardInterrupt:
    print("Shutting down...")
finally:
    hri_connector.shutdown()
    rclpy.shutdown()

After a while, you should hear speech output from TTSAgent.

With ROS2S2SAgent

Run the following script and converse with agent:

from rai_s2s.sound_device import SoundDeviceConfig
from rai.communication.ros2 import ROS2Context
from rai_s2s.s2s.agents.s2s_agent import SpeechToSpeechAgent
from rai_s2s.s2s.agents.ros2s2s_agent import ROS2S2SAgent
from rai.agents.langchain.react_agent import ReActAgent
from rai_s2s.asr.models import OpenAIWhisper, SileroVAD
from rai_s2s import KokoroTTS

from rai.agents import AgentRunner


@ROS2Context()
def main():
    speaker_config = SoundDeviceConfig(
        stream=True,
        is_output=True,
        # device_name="EPOS PC 8 USB: Audio (hw:1,0)",
        # device_name="Sennheiser USB headset: Audio (hw:1,0)",
        # device_name="Jabra Speak2 40 MS: USB Audio (hw:2,0)",
        device_name="default",
    )

    microphone_config = SoundDeviceConfig(
        stream=True,
        channels=1,
        device_name="default",
        consumer_sampling_rate=16000,
        dtype="int16",
        is_input=True,
    )

    # whisper = LocalWhisper("tiny", 16000)
    whisper = OpenAIWhisper("gpt-4o-mini-transcribe", 16000)
    vad = SileroVAD(16000, 0.5)
    
    tts = KokoroTTS()

    agent = ROS2S2SAgent(
        from_human_topic="/from_human",
        to_human_topic="/to_human",
        microphone_config=microphone_config,
        speaker_config=speaker_config,
        transcription_model=whisper,
        vad=vad,
        tts=tts,
    )
    from rai.communication.ros2 import ROS2HRIConnector

    hri_connector = ROS2HRIConnector()
    llm = ReActAgent(
        target_connectors={"/to_human": hri_connector},
    )
    llm.subscribe_source("/from_human", hri_connector)
    runner = AgentRunner([agent, llm])
    runner.run_and_wait_for_shutdown()


if __name__ == "__main__":
    main()

The KokoroTTS model works well together with the ROS2S2SAgent.
My UX - It sounds nicer compared with OpenTTS. I didn't observe any significant differences in inference time between the models.
The model sometimes does not put space between the sentences. EDIT: It was fixed by setting trim to false in create method of Kokoro.

[maciejmajek]

Does zarr with the 3.0.6 break rai?

[maciejmajek]

Suggested change

	> [!WARNING]
	> It is not recommended to use device_name set to `'default'` in `SoundDeviceConfig` due to potential issues with audio.
	> [!TIP]
	> If you're experiencing audio issues and device_name is set to 'default', try specifying the exact device name instead, as this often resolves the problem.

[maciejmajek]

Also, please add this note to the configurator.

[maciejmajek]

Is kokoro pronouncing markdown elements?
Why does this method remove only a certain subset of markdown symbols?

[maciejmajek]

Are we expecting values outside of the provided range?
Clipping audio should only be used as a last resort, as it introduces massive quality degradation.

[maciejmajek]

Can we configure kokoro's logger to drop these warnings?