3. AI Large Model Voice Interaction

Note: Due to performance limitations, Raspberry Pi 5 2GB, 4GB, RDK X5 4GB, and Jetson Orin Nano 4GB versions cannot run offline. Please refer to online large model tutorials.

1. Concept Introduction

1.1 What is "AI Large Model Voice Interaction"?

In the largemodel project, AI Large Model Voice Interaction connects the previously introduced offline ASR and offline TTS with the core of the Large Language Model (LLM), forming a complete dialogue system that can hear, speak, and think.

This is no longer an isolated function, but a prototype of a true voice assistant. Users can engage in natural language conversations with the robot via voice, and the robot can understand questions, think about answers, and respond with voice. The entire process is completed locally, without a network.

The core of this function is the model_service ROS2 node. It acts as the brain and central nervous system, subscribing to the ASR recognition results, calling the LLM for processing, and then publishing the LLM's text response to the TTS node for speech synthesis.

1.2 Implementation Principle Overview

The implementation principle of this function is a classic data flow pipeline:

1. Audio -> Text (ASR): The asr node continuously listens to ambient sound. Once it detects that a user has finished speaking a sentence, it converts it into text and publishes it to the /asr_text topic.
2. Text -> Processing -> Text (LLM): The model_service node subscribes to the /asr_text topic. When it receives the text from the ASR, it sends it as a prompt to the locally deployed large language model (such as Qwen running via Ollama). The LLM generates a response text based on the context.
3. Text -> Audio (TTS): After receiving the LLM's response, the model_service node publishes it to the /tts_text topic.
4. Audio Playback: The tts_only node subscribes to the /tts_text topic. Upon receiving the text, it immediately calls the offline TTS model to synthesize it into audio and plays it through the speaker.

This process forms a complete closed loop of speech input -> text processing -> text output -> speech output.

2. Code Analysis

Key Code

1. Speech Input Node (largemodel/largemodel/asr.py)

xxxxxxxxxx
# From largemodel/largemodel/asr.py
class ASRNode(Node):
    def __init__(self):
        # ...
        self.asr_pub = self.create_publisher(String, "asr", 5)
        # ...
​
    def kws_handler(self)->None:
        if self.listen_for_speech(self.mic_index):
            asr_text = self.ASR_conversion(self.user_speechdir)
            if asr_text != 'error':
                self.asr_pub_result(asr_text)
​
    def asr_pub_result(self,asr_result:str)->None:
        msg=String(data=asr_result)
        self.asr_pub.publish(msg)

2. AI service and voice output node (largemodel/largemodel/model_service.py)

xxxxxxxxxx
# From largemodel/largemodel/model_service.py
class LargeModelService(Node):
    def __init__(self):
        # ...
        self.asrsub = self.create_subscription(String,'asr', self.asr_callback,1)
        # ...
​
    def asr_callback(self,msg):
        """Callback function for handling ASR messages."""
        # ...
        result = self.model_client.infer_with_text(msg.data, message=messages_to_use)
        self.process_model_result(result)
​
    def process_model_result(self, result, from_seewhat=False):
        """Process the result returned by the model. """
        # ...
        user_friendly_response = response_json.get("response", "我正在处理...")
        # ...
        self._safe_play_audio(user_friendly_response)
        # ...
        self.execute_tools(tools_list)
​
    def _safe_play_audio(self, text_to_speak: str):
        """
        Synthesizes and plays all non-empty messages only in non-text chat mode.
        """
        if not self.text_chat_mode and text_to_speak:
            try:
                self.model_client.voice_synthesis(text_to_speak, self.tts_out_path)
                self.play_audio_async(self.tts_out_path)
            except Exception as e:
                self.get_logger().error(f"Safe audio playback failed: {e}")

Code Analysis

The voice interaction functionality of the large AI model is implemented by two independent ROS nodes, asr.py and model_service.py, which communicate via the ROS topic /asr, forming a complete processing loop.

1. Voice Input and Publishing (asr.py):

   • The ASRNode node processes voice input. In the kws_handler function, it performs audio recording and text conversion.
   • After successful conversion, the asr_pub_result function is called. This function places the resulting text string into a std_msgs.msg.String message and then publishes it to the /asr topic via the self.asr_pub publisher.

2. Subscription and Response (model_service.py):

   • During initialization, the LargeModelService node subscribes to the /asr topic using the create_subscription method and specifies asr_callback as its callback function.
   • When ASRNode publishes a message on the /asr topic, the ROS system automatically calls the asr_callback function of LargeModelService, passing the message as the msg parameter.

3. Interaction Loop:

   • After receiving the message, the asr_callback function extracts the text data (msg.data) and passes it to the large model for inference (self.model_client.infer_with_text).
   • After inference is complete, the process_model_result function is called. This function extracts the reply text from the model result.
   • Next, the _safe_play_audio function is called, which uses the TTS function (self.model_client.voice_synthesis) to synthesize the reply text into audio and play it back.
   • At this point, a complete interaction loop from "receiving user voice" to "responding to the user with voice" is completed. The two nodes are decoupled using ROS's publish-subscribe model. asr.py handles input, while model_service.py handles processing and output.

3. Practical Operation

3.1 Configuring Offline Voice Interaction

To implement a fully offline voice interaction system, you need to ensure that ASR, TTS, and LLM are all configured in offline mode.

Raspberry Pi 5 requires entering a Docker container; RDK X5 and Orin controllers do not require this:

xxxxxxxxxx
./ros2_docker.sh

If you need to enter the same Docker container later and run other commands, simply enter ./ros2_docker.sh again in the host terminal.

1. Open the main configuration file yahboom.yaml:

   xxxxxxxxxx
   vim ~/yahboom_ws/src/largemodel/config/yahboom.yaml
   

2. Modify/confirm the following key configurations:

   xxxxxxxxxx
   asr:
     ros__parameters:
       use_oline_asr: False                # Key: Set to False to enable offline ASR
       regional_setting : "China" 
   ​
   model_service:
     ros__parameters:
       useolinetts: False                   # Key: Set to False to enable offline TTS
       llm_platform: 'ollama'              # Key: Set to 'ollama' to enable offline LLM
   regional_setting : "China" 
   

3. Open the model interface configuration file large_model_interface.yaml:

   xxxxxxxxxx
   vim ~/yahboom_ws/src/largemodel/config/large_model_interface.yaml
   

4. Confirm that all offline model paths are correct:

   • Raspberry Pi 5

   xxxxxxxxxx
   # large_model_interface.yaml
   ​
   ## Offline large model
   ollama_model: "qwen2.5:7b"  # Make sure the model has been downloaded via ollama pull
   ​
   ## Offline speech recognition
   local_asr_model: "/path/to/your/SenseVoiceSmall" # Ensure the ASR model path is correct
   ​
   ## Offline speech synthesis
   zh_tts_model: "/path/to/your/zh_CN-huayan-medium.onnx" # 确保TTS模型路径正确
   # ...
   

   • RDK X5

   xxxxxxxxxx
   # large_model_interface.yaml
   ​
   ## Offline large model
   ollama_model: "qwen2.5:7b"  # Make sure the model has been downloaded via ollama pull
   ​
   ## Offline speech recognition
   local_asr_model: "~/yahboom_ws/src/largemodel/MODELS/asr/SenseVoiceSmall" # Ensure the ASR model path is correct
   ​
   ##Offline speech synthesis (Offline TTS)
   # Chinese TTS model 
   zh_tts_model: "/home/sunrise/yahboom_ws/src/largemodel/MODELS/tts/zh/zh_CN-huayan-medium.onnx"
   zh_tts_json: "/home/sunrise/yahboom_ws/src/largemodel/MODELS/tts/zh/zh_CN-huayan-medium.onnx.json"
    Offline Text-to-Speech (TTS)
   Multimodal visual understanding and voice interaction