wtf is robotlegs-stt?

RobotLegs STT is a tool built in the Rust programming language for converting spoken English audio into written text. It uses a specific artificial intelligence model from NVIDIA called Canary-Qwen 2.5B. The project provides both a software library that developers can include in their own applications and a command-line program for direct use. The main goal is to perform speech recognition efficiently on specific NVIDIA hardware. The software is designed with a strict hardware requirement. It works only on NVIDIA GPUs with a compute capability of 12.0, such as the RTX 5090 graphics card. If you try to run it on any other type of GPU, the software will refuse to operate. This narrow focus allows the developers to optimize the system heavily. The system processes audio in stages: first converting raw sound into numerical features, then turning those features into speech embeddings, and finally generating text from those embeddings. To handle the large model files and runtime components, RobotLegs downloads a pinned bundle of eight files totaling about 3.7 GB. Every file is checked for its size and a cryptographic hash before it is used, ensuring the files have not been altered. The software manages these downloads carefully, making them resumable and verifying each file before placing it into active use. This means developers do not need to manually configure environment variables or worry about where the libraries are stored. The audio input must meet specific criteria. The software accepts WAV or native FLAC files, and the audio must be mono, recorded at exactly 16 kHz, and between 20 milliseconds and 40 seconds long. If you have audio in a different format, you must convert it first. The developers provide an example using a tool called ffmpeg to convert an Opus audio file into the required FLAC format. The developers measured performance on a certified RTX 5090 host using a 13.69-second audio clip. Compared to the official NVIDIA reference implementation, the production setup achieved higher throughput, lower latency, and less GPU memory usage. Both the reference system and RobotLegs produced consistent results across 30 test runs, with only a minor difference in how one word was transcribed. The full README is longer than what was shown.