Issues with TTS and ASR on Nvidia Jetson Orin Nano Dev Board

Issue Overview

Users are experiencing difficulties getting Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) functionalities to work on the Nvidia Jetson Orin Nano 8GB development board. The main issues include:

1. Inability to run TTS or ASR-related tasks
2. Problems with the Jetpack and NVIDIA ecosystem
3. Difficulties installing and running various software packages, including Riva, NeMo, and pynini

Users have reported feeling frustrated with the overall experience, questioning the value of their investment in the device.

Possible Causes

1. Incompatible software versions: The Jetson Orin Nano may require specific versions of software packages that are not readily available or compatible with the current system.

2. Outdated or incorrect installation procedures: The installation methods provided in documentation may not be up-to-date or suitable for the Orin Nano.

3. Hardware limitations: The Orin Nano’s specifications may not fully support certain TTS and ASR functionalities out of the box.

4. Dependency conflicts: There may be conflicts between different software packages and their dependencies, causing installation and runtime errors.

5. Insufficient system resources: The device may lack adequate memory or processing power to run certain TTS and ASR tasks efficiently.

Troubleshooting Steps, Solutions & Fixes

1. Update Jetpack and NVIDIA ecosystem:

   • Use the NVIDIA SDK Manager to flash the latest compatible Jetpack version for the Orin Nano.
   • Ensure you’re using a native Ubuntu machine to run the SDK Manager for proper NVMe installation.

2. Install Riva correctly:

   • Use NGC CLI version 3.22 instead of the latest version (3.23) due to known issues.
   • Download Riva quickstart files using the command:

     ngc registry resource download-version "nvidia/riva/riva_quickstart_arm64:2.11.0"
     

3. Install OpenFST and pynini:

   • Build and install OpenFST 1.8.2 from source, as pynini requires a newer version than the one provided by Ubuntu.
   • After installing OpenFST, attempt to install pynini and nemo_toolkit.

4. Use the l4t-ml container:

   • Clone the jetson-containers repository:

     git clone https://github.com/dusty-nv/jetson-containers
     

   • Build the l4t-ml container, which includes the necessary setup for NeMo and other AI tools.

5. Manage system resources:

   • If encountering "Killed" messages during compilation, it may indicate insufficient memory.
   • Disable zram and mount additional swap space:

     sudo systemctl disable nvzramconfig
     sudo fallocate -l 8G /mnt/8GB.swap
     sudo mkswap /mnt/8GB.swap
     sudo swapon /mnt/8GB.swap
     

   • Add the following line to /etc/fstab to make the swap permanent:

     /mnt/8GB.swap  none  swap  sw 0  0
     

   • If needed, disable the desktop UI to free up resources.

6. Adjust build parameters:

   • If compilation still fails, try adding the following line to the Dockerfile.pytorch:

     export NINJA_FLAGS=-j 1 &&
     

   • This limits the number of parallel jobs during compilation, potentially reducing memory usage.

7. Use ROS wrappers:

   • For integrating Riva with ROS, create custom ROS wrapper nodes or model them after existing ones in the jetson-voice project.

8. Community support:

   • Stay updated with the latest developments and solutions by following the Nvidia forums and GitHub repositories related to Jetson development.
   • Contribute to the community by reporting issues and sharing successful configurations.

By following these steps and continuously updating your system with the latest compatible software versions, you should be able to improve the functionality of TTS and ASR on your Nvidia Jetson Orin Nano Dev Board. Remember that this is an evolving platform, and new solutions may become available as the community and Nvidia continue to address these issues.