Melba-stt

A rust discord bot that joins a voice channel and transcribes spoken audio from each user.

Running

1. Install the rust toolchain

With CUDA and cuDNN (recommended)

2. Install CUDA and cuDNN
3. Run cargo run --release

Cpu only

2. Run cargo run --release --no-default-features

Furthermore, you may experience a performance improvement if you compile for the native processor (as simd instructions may be used). This can be activated by using RUSTFLAGS="-C target-cpu=native" cargo run --release.

Configuration

The bot is configured using a bot.toml file in the directory that the bot is ran. An example of the values that can go in this file are located in bot.example.toml. Furthermore, the token of the bot is located in secrets.toml and has a corresponding example file in secrets.example.toml. By default, the bot will use openai's tiny model for whisper, which is equivalent to having the following line in the config file:

model = { repo = { name = "openai/whisper-tiny" } }

If you want to use another model, then the repo can be changed.

License

(c) by NOM Network

Code is licensed under an AGPL 3.0 (and later) license.