This project provides low-resource streaming speech recognition using the Picovoice Cheetah system (free). It processes speech offline and is available as a C++ class or a Python wrapper. The code is known to work on Jetson Nano running Ubuntu 18.04 and Raspberry Pi 4 running Bullseye (but not Buster). Its latency is under a second, as seen in this video.
All speech commands (as delimited by pauses) are transcribed with no special attention word. To use the system call the "start" function, loop using the "status" function, then pull result strings using the "heard" function. The simple program test_reco shows how to do this (or see __main__ in the Python wrapper). Below is an example session:
Configuring Picovoice and connecting to microphone ...
--- Transcribing utterances for 20 secs ---
What ...
What color is the biggest ...
What color is the biggest object on the table in the kitchen
--- Done ---
The system runs in a background thread and uses the Linux pulseaudio front end (sorry, no Windows) to access the default microphone. The code is pretty vanilla except for the inclusion of a rough VOX circuit which skips speech recognition when things are quiet. This lets the system catch up on any lags induced by the highly variable delays in audio sampling and Picovoice calls.
This code needs a valid Picovoice key to operate! You can sign up for free then copy your AccessKey string to file picovoice.key. Unfortunately, your robot must have an active internet connection in order to validate and run with this key, thus negating a prime advantange. Maybe Picovoice will offer a local license in the future ...
The code here has been compiled for RPi4 64 bit and should be usable as is. If you want to build for Jetson Nano instead, change which library is commented out on line 43 of CMakeLists.txt. To run the code you will also need pulseaudio installed:
sudo apt install pulseaudio libasound2
Finally, this repository has release v2.0 of the Picovoice Cheetah shared library and speech model. The newest libraries for RPi are Cortex A72 32 bit and Cortex A72 64 bit, and for Jetson it is Cortex A57 64 bit. You might also want to check for an updated version of the run-time model cheetah_params on GitHub.
February 2024 - Jonathan Connell - [email protected]