Simple Offline Speech Recognition (STT)

SimpleOfflineSTT is a package that will let you perform speech recognition on AudioClips, raw audio data, or microphone input using Unity Inference. It's completely offline without any networking or services required. You can try it out in a WebGPU demo here.

It does not have any platform-specific code or libraries, so can be used on any platform supported by the Inference package (Windows, Mac, WebGL and Android tested). Inference can be performed on the CPU or GPU.

For transcription, OpenAI's pre-trained Whisper-Tiny speech-to-text model is used. It is very accurate, and supports over 90 languages. You can use the base models (~ 400MB total), the fp16 quantized versions (~200MB total) or the uint8 versions (~105MB total). They all have good accuracy.

Whisper-Tiny is distributed using the Apache v2.0 licence, which makes it usable in commercial projects, and can be redistributed with minimal restrictions. Please see the included licence file for more information.