随着人工智能技术的快速发展，语音交互已成为越来越重要的人机交互模式。特别是在智能家居、个人助理和客户服务支持等领域，对语音聊天机器人的需求正在显着增长。然而，现有的大多数语音聊天机器人都依赖于云计算服务，这在一定程度上引发了对数据隐私和网络延迟的担忧.
该项目旨在通过构建本地运营的语音聊天机器人来解决这些问题。利用 Nvidia Riva 和 Meta Llama2，我们开发了一个安全、私密且快速响应的语音交互系统。
材料清单
- 具有超过 16GB 内存的 Jetson 设备。
- 硬件设备需要使用 jetpack 5.1.1 操作系统进行预刷机。
- 扬声器和麦克风。
注意：我使用 Jetson AGX Orin 32GB H01 套件完成了所有实验，但您可以尝试使用内存较少的设备加载较小的模型
制作步骤
硬件连接
将音频输入/输出设备连接到 reComputer。
打开重新计算机的电源，并确保其具有正常的网络访问权限。
安装 Riva 服务器，详情请参考此处：
https://docs.nvidia.com/deepl...
第1步，访问并登录 NVIDIA NGC
https://catalog.ngc.nvidia.co...
第2步，获取 NGC API 密钥
Account(top right corner) --> Setup --> Get API Key --> Generate API Key --> Confirm
第3步，在 reComputer 上配置 NGC
打开reComputer终端（您可以使用快捷键在reComputer的桌面上快速打开终端，也可以使用另一台计算机远程访问reComputer的终端）并逐个输入以下命令。Ctrl+Alt+T
cd ~ && mkdir ngc_setup && cd ngc_setup
wget --content-disposition https://api.ngc.nvidia.com/v2/resources/nvidia/ngc-apps/ngc_cli/versions/3.36.0/files/ngccli_arm64.zip && unzip ngccli_arm64.zip
chmod u+x ngc-cli/ngc
echo "export PATH=\"\$PATH:$(pwd)/ngc-cli\"" >> ~/.bash_profile && source ~/.bash_profile
ngc config set在终端交互界面输入相关信息。
第4步，在reComputer上安装并运行Riva服务器
在reComputer的终端中，输入以下命令。
cd ~ && mkdir riva_setup && cd riva_setup
ngc registry resource download-version nvidia/riva/riva_quickstart_arm64:2.13.1
cd riva_quickstart_v2.13.1用于修改配置Vim文件。
vim config.sh
# Press the `A` key on the keyboard to enter edit mode.
# Edit lines 18 and 20 following the instructions below.
# service_enabled_nlp=true --> service_enabled_nlp=false
# service_enabled_nmt=true --> service_enabled_nmt=false
# Press the `ESC` on the keyboard to exit edit mode, then use the shortcut `Shift+Z Z` to save the edited content and close the editor.编辑后的 Config.sh 配置文件：
# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto. Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.
# GPU family of target platform. Supported values: tegra, non-tegra
riva_target_gpu_family="non-tegra"
# Name of tegra platform that is being used. Supported tegra platforms: orin, xavier
riva_tegra_platform="orin"
# Enable or Disable Riva Services
# For any language other than en-US: service_enabled_nlp must be set to false
service_enabled_asr=true
service_enabled_nlp=false
service_enabled_tts=true
service_enabled_nmt=false
# Configure translation services
# Text-to-Text translation (T2T):
# - service_enabled_nmt must be set to true
# - Uncomment desired model for source and target languages in models_nmt field
# Speech-to-Text translation (S2T):
# - service_enabled_asr, service_enabled_nmt must be set to true
# - Set language code of input speech in the asr_language_code field
# - Uncomment desired model for source and target languages in models_nmt field
# Speech-to-Speech translation (S2S):
# - service_enabled_asr, service_enabled_nmt, service_enabled_tts must be set to true
# - Set language code of input speech in the asr_language_code field
# - Uncomment desired model for source and target languages in models_nmt field
# - Set language code of output speech in the tts_language_code field
# Enable Riva Enterprise
# If enrolled in Enterprise, enable Riva Enterprise by setting configuration
# here. You must explicitly acknowledge you have read and agree to the EULA.
# RIVA_API_KEY=<ngc api key>
# RIVA_API_NGC_ORG=<ngc organization>
# RIVA_EULA=accept
# Language code to fetch ASR models of a specific language
# Supported language codes: ar-AR, en-US, en-GB, de-DE, es-ES, es-US, fr-FR, hi-IN, it-IT, ja-JP, ru-RU, ko-KR, pt-BR, zh-CN, es-en-US, ja-en-JP
# For multiple languages enter space separated language codes.
asr_language_code=("en-US")
# ASR acoustic model architecture
# Supported values are: conformer, conformer_xl (en-US + amd64 only), citrinet_1024, citrinet_256 (en-US + arm64 only), jasper (en-US + amd64 only), quartznet (en-US + amd64 only)
asr_acoustic_model=("conformer")
# ASR acoustic model architecture variant
# Supported values for the architecture are:
# conformer: unified(de-DE, ja-JP and zh-CN only), ml_cs(es-en-US only), unified_ml_cs(ja-en-JP only)
# For the default model, keep the field empty
asr_acoustic_model_variant=("")
# ASR decoder type to be used
# If you'd like to use greedy decoder for ASR instead of flashlight/os2s decoder then set the below $use_asr_greedy_decoder to true
use_asr_greedy_decoder=false
# Language code to fetch TTS models of a specific language
# Supported language codes: en-US, es-ES, it-IT, de-DE, zh-CN
# For multiple languages enter space separated language codes
tts_language_code=("en-US")
# Specify one or more GPUs to use
# specifying more than one GPU is currently an experimental feature, and may result in undefined behaviours.
gpus_to_use="device=0"
# Specify the encryption key to use to deploy models
MODEL_DEPLOY_KEY="tlt_encode"
# Locations to use for storing models artifacts
#
# If an absolute path is specified, the data will be written to that location
# Otherwise, a Docker volume will be used (default).
#
# riva_init.sh will create a `rmir` and `models` directory in the volume or
# path specified.
#
# RMIR ($riva_model_loc/rmir)
# Riva uses an intermediate representation (RMIR) for models
# that are ready to deploy but not yet fully optimized for deployment. Pretrained
# versions can be obtained from NGC (by specifying NGC models below) and will be
# downloaded to $riva_model_loc/rmir by `riva_init.sh`
#
# Custom models produced by NeMo or TLT and prepared using riva-build
# may also be copied manually to this location $(riva_model_loc/rmir).
#
# Models ($riva_model_loc/models)
# During the riva_init process, the RMIR files in $riva_model_loc/rmir
# are inspected and optimized for deployment. The optimized versions are
# stored in $riva_model_loc/models. The riva server exclusively uses these
# optimized versions.
riva_model_loc="riva-model-repo"
if [[ $riva_target_gpu_family == "tegra" ]]; then
riva_model_loc="`pwd`/model_repository"
fi
# The default RMIRs are downloaded from NGC by default in the above $riva_rmir_loc directory
# If you'd like to skip the download from NGC and use the existing RMIRs in the $riva_rmir_loc
# then set the below $use_existing_rmirs flag to true. You can also deploy your set of custom
# RMIRs by keeping them in the riva_rmir_loc dir and use this quickstart script with the
# below flag to deploy them all together.
use_existing_rmirs=false
# Ports to expose for Riva services
riva_speech_api_port="50051"
# NGC orgs
riva_ngc_org="nvidia"
riva_ngc_team="riva"
riva_ngc_image_version="2.13.1"
riva_ngc_model_version="2.13.0"
# Pre-built models listed below will be downloaded from NGC. If models already exist in $riva-rmir
# then models can be commented out to skip download from NGC
########## ASR MODELS ##########
models_asr=()
for lang_code in ${asr_language_code[@]}; do
modified_lang_code="${lang_code//-/_}"
modified_lang_code=${modified_lang_code,,}
decoder=""
if [ "$use_asr_greedy_decoder" = true ]; then
decoder="_gre"
fi
if [[ ${asr_acoustic_model_variant} != "" ]]; then
if [[ ${asr_acoustic_model} == "conformer" && ${asr_acoustic_model_variant} != "unified" && ${asr_acoustic_model_variant} != "ml_cs" && ${asr_acoustic_model_variant} != "unified_ml_cs" ]]; then
echo "Valid variants for Conformer are: unified, ml_cs and unified_ml_cs."
exit 1
elif [[ ${asr_acoustic_model} != "conformer" ]]; then
echo "Invalid variant for ${asr_acoustic_model}."
exit 1
fi
asr_acoustic_model_variant="_${asr_acoustic_model_variant}"
fi
if [[ ${asr_acoustic_model} == "conformer_xl" && ${lang_code} != "en-US" ]]; then
echo "Conformer-XL acoustic model is only available for language code en-US."
exit 1
fi
if [[ ${asr_acoustic_model_variant} == "_unified" && ${lang_code} != "de-DE" && ${lang_code} != "ja-JP" && ${lang_code} != "zh-CN" ]]; then
echo "Unified Conformer acoustic model is only available for language code de-DE, ja-JP and zh-CN."
exit 1
fi
if [[ ${asr_acoustic_model_variant} == "_ml_cs" && ${lang_code} != "es-en-US" ]]; then
echo "Multilingual Code Switch Conformer acoustic model is only available for language code es-en-US."
exit 1
fi
if [[ ${asr_acoustic_model_variant} == "_unified_ml_cs" && ${lang_code} != "ja-en-JP" ]]; then
echo "Unified Multilingual Code Switch Conformer acoustic model is only available for language code ja-en-JP."
exit 1
fi
if [[ $riva_target_gpu_family == "tegra" ]]; then
if [[ ${asr_acoustic_model} == "jasper" || \
${asr_acoustic_model} == "quartznet" || \
${asr_acoustic_model} == "conformer_xl" ]]; then
echo "Conformer-XL, Jasper and Quartznet models are not available for arm64 architecture"
exit 1
fi
if [[ ${asr_acoustic_model} == "citrinet_256" && ${lang_code} != "en-US" ]]; then
echo "For arm64 architecture, citrinet_256 acoustic model is only available for language code en-US."
exit 1
fi
models_asr+=(
### Streaming w/ CPU decoder, best latency configuration
"${riva_ngc_org}/${riva_ngc_team}/models_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_str:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
### Offline w/ CPU decoder
# "${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_ofl${decoder}:${riva_ngc_model_version}"
)
else
if [[ ${asr_acoustic_model} != "conformer" && \
${asr_acoustic_model} != "conformer_xl" && \
${asr_acoustic_model} != "citrinet_1024" && \
${asr_acoustic_model} != "jasper" && \
${asr_acoustic_model} != "quartznet" ]]; then
echo "For amd64 architecture, valid acoustic models are conformer, conformer_xl, citrinet_1024, jasper and quartznet."
exit 1
fi
if [[ (${asr_acoustic_model} == "jasper" || \
${asr_acoustic_model} == "quartznet") && \
${lang_code} != "en-US" ]]; then
echo "jasper and quartznet acoustic models are only available for language code en-US."
exit 1
fi
models_asr+=(
### Streaming w/ CPU decoder, best latency configuration
"${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_str${decoder}:${riva_ngc_model_version}"
### Streaming w/ CPU decoder, best throughput configuration
# "${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_str_thr${decoder}:${riva_ngc_model_version}"
### Offline w/ CPU decoder
"${riva_ngc_org}/${riva_ngc_team}/rmir_asr_${asr_acoustic_model}${asr_acoustic_model_variant}_${modified_lang_code}_ofl${decoder}:${riva_ngc_model_version}"
)
fi
### Punctuation model
if [[ ${asr_acoustic_model_variant} != "_unified" && ${asr_acoustic_model_variant} != "_unified_ml_cs" ]]; then
pnc_lang=$(echo $modified_lang_code | cut -d "_" -f 1)
pnc_region=${modified_lang_code##*_}
modified_lang_code=${pnc_lang}_${pnc_region}
if [[ $riva_target_gpu_family == "tegra" ]]; then
models_asr+=(
"${riva_ngc_org}/${riva_ngc_team}/models_nlp_punctuation_bert_base_${modified_lang_code}:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
else
models_asr+=(
"${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_punctuation_bert_base_${modified_lang_code}:${riva_ngc_model_version}"
)
fi
fi
done
### Speaker diarization model
models_asr+=(
# "${riva_ngc_org}/${riva_ngc_team}/rmir_diarizer_offline:${riva_ngc_model_version}"
)
########## NLP MODELS ##########
if [[ $riva_target_gpu_family == "tegra" ]]; then
models_nlp=(
### Bert base Punctuation model
"${riva_ngc_org}/${riva_ngc_team}/models_nlp_punctuation_bert_base_en_us:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
### BERT Base Intent Slot model for misty domain fine-tuned on weather, smalltalk/personality, poi/map datasets.
# "${riva_ngc_org}/${riva_ngc_team}/models_nlp_intent_slot_misty_bert_base:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
### DistilBERT Intent Slot model for misty domain fine-tuned on weather, smalltalk/personality, poi/map datasets.
# "${riva_ngc_org}/${riva_ngc_team}/models_nlp_intent_slot_misty_distilbert:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
else
models_nlp=(
### Bert base Punctuation model
"${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_punctuation_bert_base_en_us:${riva_ngc_model_version}"
### BERT base Named Entity Recognition model fine-tuned on GMB dataset with class labels LOC, PER, ORG etc.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_named_entity_recognition_bert_base:${riva_ngc_model_version}"
### BERT Base Intent Slot model fine-tuned on weather dataset.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_intent_slot_bert_base:${riva_ngc_model_version}"
### BERT Base Question Answering model fine-tuned on Squad v2.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_question_answering_bert_base:${riva_ngc_model_version}"
### Megatron345M Question Answering model fine-tuned on Squad v2.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_question_answering_megatron:${riva_ngc_model_version}"
### Bert base Text Classification model fine-tuned on 4class (weather, meteorology, personality, nomatch) domain model.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_nlp_text_classification_bert_base:${riva_ngc_model_version}"
)
fi
########## TTS MODELS ##########
models_tts=()
for lang_code in ${tts_language_code[@]}; do
modified_lang_code="${lang_code//-/_}"
modified_lang_code=${modified_lang_code,,}
if [[ $riva_target_gpu_family == "tegra" ]]; then
if [[ ${lang_code} == "en-US" ]]; then
models_tts+=(
### These models have been trained with energy conditioning and use the International Phonetic Alphabet (IPA) for inference and training.
"${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_en_us_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
# "${riva_ngc_org}/${riva_ngc_team}/models_tts_radtts_hifigan_en_us_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
### This model uses the ARPABET for inference and training.
# "${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_en_us:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
elif [[ ${lang_code} == "zh-CN" ]]; then
models_tts+=(
### This model is multi-speaker with emotion and and use the International Phonetic Alphabet (IPA) for inference and training.
"${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_zh_cn_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
else
### These models are single-speaker and use the International Phonetic Alphabet (IPA) for inference and training.
if [[ ${lang_code} != "de-DE" ]]; then
models_tts+=(
"${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_${modified_lang_code}_f_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
fi
models_tts+=(
"${riva_ngc_org}/${riva_ngc_team}/models_tts_fastpitch_hifigan_${modified_lang_code}_m_ipa:${riva_ngc_model_version}-${riva_target_gpu_family}-${riva_tegra_platform}"
)
fi
else
if [[ ${lang_code} == "en-US" ]]; then
models_tts+=(
### These models have been trained with energy conditioning and use the International Phonetic Alphabet (IPA) for inference and training.
"${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_en_us_ipa:${riva_ngc_model_version}"
# "${riva_ngc_org}/${riva_ngc_team}/rmir_tts_radtts_hifigan_en_us_ipa:${riva_ngc_model_version}"
### This model uses the ARPABET for inference and training.
# "${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_en_us:${riva_ngc_model_version}"
)
elif [[ ${lang_code} == "zh-CN" ]]; then
models_tts+=(
### This model is multi-speaker with emotion and and use the International Phonetic Alphabet (IPA) for inference and training.
"${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_zh_cn_ipa:${riva_ngc_model_version}"
)
else
### These models are single-speaker and use the International Phonetic Alphabet (IPA) for inference and training.
if [[ ${lang_code} != "de-DE" ]]; then
models_tts+=(
"${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_${modified_lang_code}_f_ipa:${riva_ngc_model_version}"
)
fi
models_tts+=(
"${riva_ngc_org}/${riva_ngc_team}/rmir_tts_fastpitch_hifigan_${modified_lang_code}_m_ipa:${riva_ngc_model_version}"
)
fi
fi
done
######### NMT models ###############
# Models follow Source language _ One or more target languages model architecture
# Source or target language "any" means the model supports 32 languages mentioned in docs.
# e.g., rmir_nmt_de_en_24x6 is a German to English 24x6 bilingual model
# and rmir_megatronnmt_en_any_500m is a English to 32 languages megatron model
models_nmt=(
###### Bilingual models
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_de_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_es_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_zh_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_ru_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_fr_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_de_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_es_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_ru_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_zh_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_fr_en_24x6:${riva_ngc_model_version}"
###### Multilingual models
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_deesfr_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_en_deesfr_12x2:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_deesfr_en_24x6:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_nmt_deesfr_en_12x2:${riva_ngc_model_version}"
###### Megatron models
#"${riva_ngc_org}/${riva_ngc_team}/rmir_megatronnmt_any_en_500m:${riva_ngc_model_version}"
#"${riva_ngc_org}/${riva_ngc_team}/rmir_megatronnmt_en_any_500m:${riva_ngc_model_version}"
)
NGC_TARGET=${riva_ngc_org}
if [[ ! -z ${riva_ngc_team} ]]; then
NGC_TARGET="${NGC_TARGET}/${riva_ngc_team}"
else
team="\"\""
fi
# Specify paths to SSL Key and Certificate files to use TLS/SSL Credentials for a secured connection.
# If either are empty, an insecure connection will be used.
# Stored within container at /ssl/servert.crt and /ssl/server.key
# Optional, one can also specify a root certificate, stored within container at /ssl/root_server.crt
ssl_server_cert=""
ssl_server_key=""
ssl_root_cert=""
# define Docker images required to run Riva
image_speech_api="nvcr.io/${NGC_TARGET}/riva-speech:${riva_ngc_image_version}"
# define Docker images required to setup Riva
image_init_speech="nvcr.io/${NGC_TARGET}/riva-speech:${riva_ngc_image_version}-servicemaker"
# daemon names
riva_daemon_speech="riva-speech"
if [[ $riva_target_gpu_family != "tegra" ]]; then
riva_daemon_client="riva-client"
Fi 使用类似的方法修改 /etc/docker/daemon.json。
sudo vim /etc/docker/daemon.json
# Add this line >> "default-runtime": "nvidia"
# Press the `ESC` on the keyboard to exit edit mode, then use the shortcut `Shift+Z Z` to save the edited content and close the editor.
sudo systemctl restart docker编辑后的配置文件如下：
/etc/docker/daemon.json
{
"default-runtime": "nvidia",
"runtimes": {
"nvidia": {
"path": "nvidia-container-runtime",
"runtimeArgs": []
}
}
}
使用以下命令初始化并启动 Riva：
sudo bash riva_init.sh
sudo bash riva_start.sh
安装并运行 LLM
为了简化安装过程，我们可以参考 Dusty 的 jetson-containers 项目来安装文本生成推理，并使用文本生成推理来加载 Llama2 7B 大型语言模型。打开一个新终端并运行以下命令。
cd ~
git clone https://github.com/dusty-nv/jetson-containers.git
cd jetson-containers
pip install -r requirements.txt
./run.sh $(./autotag text-generation-inference)
export HUGGING_FACE_HUB_TOKEN=<your hugging face token>
text-generation-launcher --model-id meta-llama/Llama-2-7b-chat-hf --port 8899
克隆本地聊天机器人演示并尝试运行它。
现在，您应该至少打开两个终端，一个运行 Riva 服务器，另一个运行文本生成推理服务器。接下来，我们需要打开一个新终端来运行我们的演示。
cd ~
git clone https://github.com/yuyoujiang/Deploy-Riva-LLama-on-Jetson.git
cd Deploy-Riva-LLama-on-Jetson
# Query audio input/output devices.
python3 local_chatbot.py --list-input-devices
python3 local_chatbot.py --list-output-devices
python3 local_chatbot.py --input-device <your device id> --output-device <your device id>
# For example: python3 local_chatbot.py --input-device 25 --output-device 30
