Struct nnnoiseless::DenoiseState

pub struct DenoiseState<'model> { /* fields omitted */ }

This is the main entry-point into nnnoiseless. It mainly contains the various memory buffers that are used while denoising. As such, this is quite a large struct, and should probably be kept behind some kind of pointer.

Example

// One second of 440Hz sine wave at 48kHz sample rate. Note that the input data consists of
// `f32`s, but the values should be in the range of an `i16`.
let sine: Vec<_> = (0..48_000)
    .map(|x| (x as f32 * 440.0 * 2.0 * std::f32::consts::PI / 48_000.0).sin() * i16::MAX as f32)
    .collect();
let mut output = Vec::new();
let mut out_buf = [0.0; DenoiseState::FRAME_SIZE];
let mut denoise = DenoiseState::new();
let mut first = true;
for chunk in sine.chunks_exact(DenoiseState::FRAME_SIZE) {
    denoise.process_frame(&mut out_buf[..], chunk);

    // We throw away the first output, as discussed in the documentation for
    //`DenoiseState::process_frame`.
    if !first {
        output.extend_from_slice(&out_buf[..]);
    }
    first = false;
}

Implementations

impl DenoiseState<'static>

pub const FRAME_SIZE: usize

A DenoiseState processes this many samples at a time.

pub fn new() -> Box<DenoiseState<'static>>

Creates a new DenoiseState.

pub fn from_model(model: RnnModel) -> Box<DenoiseState<'static>>

Creates a new DenoiseState owning a custom model.

The main difference between this method and DenoiseState::with_model is that here DenoiseState will own the model; this might be more convenient.

impl<'model> DenoiseState<'model>

pub fn with_model(model: &'model RnnModel) -> Box<DenoiseState<'model>>

Creates a new DenoiseState using a custom model.

The main difference between this method and DenoiseState::from_model is that here DenoiseState will borrow the model; this might create some lifetime-related pain, but it means that the same model can be shared between multiple DenoiseStates.

pub fn process_frame(&mut self, output: &mut [f32], input: &[f32]) -> f32

Processes a chunk of samples.

Both output and input should be slices of length DenoiseState::FRAME_SIZE, and they are assumed to be in 16-bit, 48kHz signed PCM format. Note that although the input and output are f32s, they are supposed to come from 16-bit integers. In particular, they should be in the range [-32768.0, 32767.0] instead of the range [-1.0, 1.0] which is more common for floating-point PCM.

The current output of process_frame depends on the current input, but also on the preceding inputs. Because of this, you might prefer to discard the very first output; it will contain some fade-in artifacts.

Trait Implementations

impl<'model> Clone for DenoiseState<'model>

fn clone(&self) -> DenoiseState<'model>

Returns a copy of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.