Edge AI voice assistant with Raspberry Pi 5 and AI HAT+ 2

Installing the OS

Write the Raspberry Pi 5 OS to the SD card. This time, we used Raspberry Pi Imager.

Once you download the software, launch it.
Select and enter the following items:

Once you have completed all the settings, click WRITE to write the data.

Installing the Raspberry Pi AI HAT+ 2

Next, install the Raspberry Pi AI HAT+ 2 onto the Raspberry Pi 5.

The box contains the following parts:

• AI HAT+ 2 board
• I/O extension pin header
• heat sink
• Spacers, screws

First, attach the heatsink. Peel off the sheet on the back of the heatsink. Refer to the instruction card to align the pins properly. The Raspberry Pi logo on the AI HAT+ 2 board and the heatsink should be facing the same direction.

If you find it difficult to insert the pins, try placing the heat sink on the bottom and pressing the board from above. Be sure to press around the pins to avoid damaging the board.

Insert the AI HAT+ 2 board's flexible CABLES into CONNECTORS on the Raspberry Pi 5. When installing, pull up the brown slider on CONNECTORS (be careful not to apply too much force and break it!). Once inserted, push the slider down to secure CABLES.

Secure it in place with the included spacers and screws. This completes the installation.

Installing the AI HAT+ 2 (Hailo-10H) execution environment

The AI HAT+ 2 is equipped with the Hailo-10H edge AI accelerator. We will install the drivers and various tools to control this device.

1. Turn on POWER SUPPLIES and start the terminal.
A network connection is required to install the software, so connect your Raspberry Pi 5 to Wi-Fi or a wired LAN.

2. Run the following in the terminal (don't type $ as it is the terminal prompt). Note that the package name is hailo-h10-all, not 10h.

$ sudo apt update
# ドライバモジュールのビルド、インストールを行うのに必要
$ sudo apt install dkms
# AI HAT+ 2実行環境
$ sudo apt install -y hailo-h10-all
$ sudo reboot

3. After rebooting, open the terminal. Run hailortcli scan and if the following message appears, the runtime environment has been successfully installed.

$ hailortcli scan
Hailo Devices:
[-] Device: 0001:01:00.0

Obtaining the sample application code and installing the package

Reference links:

• https://github.com/hailo-ai/hailo-apps/tree/main/hailo_apps/python/gen_ai_apps/voice_assistant

1. In the terminal, go to the appropriate directory and download the source code. Once downloaded, go to the demo directory.

If you use Python for other purposes, it is recommended that you create a Python virtual environment specifically for running demos so that packages are not affected.

$ git clone https://github.com/hailo-ai/hailo-apps.git
$ cd ./hailo-apps/
# venv_gen_aiという名前のPython仮想環境を作成する
# システムにインストールされたHAILOパッケージを引き継ぐため
# --system-site-packagesを指定する
$ python -m venv --system-site-packages venv_gen_ai
# Python実行環境の切り替え
$ . venv_gen_ai/bin/activate

Note: When rebooting and running, make sure you switch virtual environments first.

2. Install the libraries and Python packages required for the demo.

$ sudo apt-get install -y portaudio19-dev
# hailo-apps/で実行
$ pip install -e ".[gen-ai]"

3. With the USB microphone and LOUDSPEAKERS connected, run the following tool. The tool will automatically recognize the connected microphone and LOUDSPEAKERS and perform a recording and playback test.

$ hailo-audio-troubleshoot

Running a Voice Assistant Demo

1. Create a directory to download the model

$ sudo mkdir /usr/local/hailo
$ sudo chown <ユーザ名>:<ユーザ名> /usr/local/hailo

2. Execute the following command to run the voice assistant demo. The first time, it will take some time to download the model.

# 初回は/usr/local/hailoにモデルをダウンロードします
$ python -m hailo_apps.python.gen_ai_apps.voice_assistant.voice_assistant

3. Once the model download is complete, the assistant will launch. Since the voice activity detection (VAD) function is not enabled, you will need to press the spacebar to indicate the start and end of VOICE INPUT UNITS.

Let's try it right away.

By default, it only recognizes English. Press the space bar to say "Hello, who are you?", and then press the space bar again when you're done. This time, the following response is returned:

Once you have confirmed that it works, enter 'Q' to exit the demo.

Japanese language support

As of the time of writing this article (2026/01/30), the voice assistant demo loads the following models:

Both models support multiple languages, so let's make it possible to input and output Japanese.

Copy the original code and create a file called voice_assistant_ja.py, then open the file in an editor.

# 音声AIアシスタントのディレクトリに移動します
$ cd hailo_apps/python/gen_ai_apps/voice_assistant
$ cp voice_assistant.py voice_assistant_ja.py
$ nano voice_assistant_ja.py

Modify the source code as follows:

Changes

    ...
    def on_audio_ready(self, audio):
      self.abort_event.clear()
  
      # 1. Transcribe
-     user_text = self.s2t.transcribe(audio)
+     # Whisperの言語コードで日本語(ja)を指定
+     user_text = self.s2t.transcribe(audio, language="ja")
  
      if not user_text:
          print("No speech detected.")
          return
      ...

Once you've made your changes, save them with Ctrl-O and exit the editor.

Now, go back to the hailo-apps directory and try running the voice assistant. Specifying the --vad option will enable automatic segment detection for VOICE INPUT UNITS, allowing it to recognize speech just by speaking into the microphone.

$ python -m hailo_apps.python.gen_ai_apps.voice_assistant.voice_assistant_ja --vad

You can now communicate in Japanese as follows:

Changes to the LLM model

The Japanese answers on Qwen2.5-1.5B-Instruct contain garbled characters and are a little unnatural.

Let's try another LLM model. Download Qwen2-1.5B-Instruct.hef from the link below to your hailo-apps directory.

• https://hailo.ai/products/hailo-software/model-explorer/generative-ai/qwen2-1-5b/

Next, open the editor and modify it to force the loading of the downloaded model.

$ nano voice_assistant_ja.py

Changes

class VoiceAssistantApp:
  ...
    def __init__(self, debug=False, no_tts=False):
        ...        
  
        if model_path is None:
            raise RuntimeError("Failed to resolve HEF path for LLM model. Please ensure the model is available.")
  
        ...
  
-       self.llm = LLM(self.vdevice, str(model_path))
+       # 読み込むモデルを指定します
+       self.llm = LLM(self.vdevice, "./Qwen2-1.5B-Instruct.hef")

Once you have made the corrections, try running the voice assistant again.

$ python -m hailo_apps.python.gen_ai_apps.voice_assistant.voice_assistant_ja --vad

The model used this time only has 1.5B parameters, so the quality of the Japanese answers is not very high. However, larger models are planned to be provided, so in the future, we will be able to communicate in Japanese more naturally.

Additionally, the time it takes for Whisper-Base to output text after you input your voice is approximately two seconds, which is quite fast for a voice assistant that runs on inexpensive edge AI.

Please try it out and use it to develop your own original applications!