NeMo-ASR : ROS Package for Speech Recognition using Nvidia NeMo

1. Intro

• This package is for offline speech recognition (ASR) using Nvidia NeMo toolkit.

Nodes

• /nemo_node : node for ASR
• exit : type 'n' for shutdown of the node.

Publishing Topics

• /speech_recognition : (String) Speech recognition result

2. Prerequisites

• Tested on python=3.8.10

Dependencies

$ apt-get install sox libsndfile1 ffmpeg portaudio19-dev
$ apt-get install build-essential

$ pip install -r requirements.txt

Nvidia NeMo

• guide @ https://github.com/NVIDIA/NeMo

$ pip install nemo_toolkit[all]

3. Usage

start NeMo node

$ roslaunch nemo_asr nemo_asr.launch \
    lang:=ko \
    frame:=5 \
    speech_channel:=speech_recognition

• lang : {"en", "ko"}
• frame : time(sec) to record each voice command
• speech_channel : topic name

Interface

[INPUT] 'y' : record for 5 seconds / 'l' : language / 'c' : cli input / 'n' : shutdown  

• press 'c' to enable command-line input (instead of STT)
• press 'l' to change language

4. Pre-trained ASR models

• this package currently uses Conformer-CTC models - https://arxiv.org/abs/2005.08100

Changing models / languages

• Currently, English and Korean is supported.
• To change the model, edit "src/utils/agent.py"

Korean ASR model

• "cwwojin/stt_kr_conformer_ctc_medium" - https://huggingface.co/cwwojin/stt_kr_conformer_ctc_medium
• This model is trained on KsponSpeech dataset - https://aihub.or.kr/
• Preprocessing & training scripts using KsponSpeech can be found at - https://github.com/rirolab/Co-op/tree/main/Woojin%20Choi/03_nemo_KsponSpeech_train

Author

• Woojin Choi / [email protected]