STT POC

This untitled project is a simple Speech-to-Text proof of concept. The project uses SDL2 to get a microphone, then dumps the data into DeepSpeech for recognoition. Using methods on Audio devices in SDL2, I was able to have SDL2 provide a data stream as raw wave i16s for DeepSpeech 0.7.4.

Dependencies

• SDL2, SDL2_ttf >= 2.0.5
• DeepSpeech 0.7.4
• Rust 1.45.0 (older versions may work, but untested)

To build

• Rust stable
• DeepSpeech Library in your path
  • I used pip install --user deepspeech==0.7.4 then added ~/.local/lib/python3.8/site-packages/deepspeech/lib to my LD_LIBRARY_PATH, which is needed for linking
• SDL2 and SDL2_ttf are needed in the library path as well, usually provided by your distribution of choice (apt, dnf, emerge, chocolatey, etc).
• Download the deepspeech model as deepspeech.pbmm and the scorer as deepspeech.scorer. and place them in the top level of the checkout, e.g. /path/to/cloned/repo/.

To use

• Update .cargo/config.toml for your target

• cargo run

• spacebar
• say something
• spacebar
• See what DeepSpeech heard
• Escape key to exit

Roadmap, In order from top -> bottom

• Move to DeepSpeech 0.8.x
• Code re-org
  • Need to get most everything out of main().
• Streaming audio instead of listen then interpret.
• VAD: Trigger words for listening instead of spacebar
• Settings
  • Font
  • Color/Style
  • Size
• Availability, produce binaries for:
  • Android
  • Linux
  • *BSD
  • MacOSX
  • Windows
  • iOS (maybe)
• More to come...