3 releases

0.0.3 Oct 6, 2024
0.0.2 Oct 4, 2024
0.0.1 Jul 23, 2024

#169 in Audio

Download history 110/week @ 2024-07-21 24/week @ 2024-07-28 14/week @ 2024-09-15 15/week @ 2024-09-22 134/week @ 2024-09-29 202/week @ 2024-10-06 20/week @ 2024-10-13 1/week @ 2024-10-20

357 downloads per month

Apache-2.0 OR MIT

115KB
2.5K SLoC

Norma

An easy to use and extensible pure Rust real-time transcription (speech-to-text) library.

Latest version Documentation License

Models

  • Whisper (with full long-form decoding support)

Exmaple

use std::{
    thread::{self, sleep},
    time::Duration,
};
use norma::{
    input::Settings,
    models::whisper::monolingual,
    Transcriber,
};

// Define the model that will be used for transcription
let model = monolingual::Definition::new(
    monolingual::ModelType::DistilLargeEnV3,
    norma::models::SelectedDevice::Cpu, // Replace with Cuda(0) or Metal as needed
);

// Spawn the transcriber in a new std thread
let (jh, th) = Transcriber::blocking_spawn(model).unwrap();

// Start recording using the default microphone
let mut stream = th.blocking_start(Settings::default()).unwrap();

thread::spawn(move || while let Some(msg) = stream.blocking_recv() {
  println!("{}", msg);
});

sleep(Duration::from_secs(10));

// Stop the transcription and drop the TranscriberHandle,
// causing the transcriber to terminate
th.stop().unwrap();
drop(th);

// Join the thread that was spawned for the transcriber
jh.join().unwrap().unwrap();

Audio backends

Norma uses cpal to be agnostic over multiple audio backends.

This allows us to support:

  • Linux (via ALSA or JACK)
  • Windows (via WASAPI)
  • macOS (via CoreAudio)
  • iOS (via CoreAudio)
  • Android (via Oboe)

Some audio backends are optional and will only be compiled with a feature flag.

  • JACK (on Linux): jack

Oboe can either use a shared or static runtime. The static runtime is used by default, but activating the oboe-shared-stdcxx feature makes it use the shared runtime, which requires libc++_shared.so from the Android NDK to be present during execution.

Accelerators

All Accelerators are defined in models::SelectedDevice.

CPU

Using the CPU does not require any extra features.

However when building on MacOS the accelerate feature can be enabled to allow the resulting program to utilize Apple's Accelerate framwork.

let device = SelectedDevice::Cpu;

CUDA and cuDNN

For the below code to compile either the cuda or the cudnn feature must be enabled.

The cuda feature flag requires that CUDA be installed and correctly configured on your machine. Once enabled the program will be built with CUDA support, and require CUDA on the machine running the code.

The cudnn feature flag requires that cuDNN be installed and correctly configured on your machine. Once enabled the program will be built with cuDNN support, and require CUDA and cuDNN on the machine running the code.

let device = SelectedDevice::Cuda(ord);

Where ord is the ID of the CUDA device you want to use. If you only have one device or want to use the default set it to 0.

Metal

Using Metal requires compiling the program on MacOS with the metal feature flag.

let device = SelectedDevice::Metal;

Dependencies

~4–46MB
~726K SLoC