Struct InputConverter 

pub(crate) struct InputConverter {
    channels: usize,
    resampling: Option<Resampling>,
}

Turns what the device delivers into the mono SAMPLE_RATE signal everything downstream reads. Pass-through when the device natively offers the model rate; otherwise downmix to mono, then resample with a windowed-sinc filter. Decimating without the filter would fold everything above the new Nyquist into the speech band as confident wrong words.

Built on the recorder thread and moved into the stream callback, which is FnMut, so push is the whole realtime path: it extends preallocated buffers and hands the samples on as they are ready.

Fields

channels: usize

resampling: Option<Resampling>

Implementations

impl InputConverter

pub(crate) fn new( channels: usize, device_rate: Option<u32>, ) -> Result<Self, Error>

channels is what the stream will deliver, device_rate the rate it will deliver at; a None rate is the pass-through.

pub(crate) fn push(&mut self, block: &[f32], out: &mut impl FnMut(&[f32], f32))

Fold one callback’s worth of device audio in, handing the mono model-rate samples and the raw block’s own peak to out.