Function speaker_from_reference_pcm 

pub fn speaker_from_reference_pcm(
    bundle: &ModelBundle,
    pcm: Vec<f32>,
    cleanup: ReferenceCleanup<'_>,
) -> Result<Vec<f32>, FttsError>

Derive a speaker vector from already-decoded mono 24 kHz reference PCM.

The PCM-level half of speaker_from_voice, split out so hosts that capture audio themselves (the iOS app through ftts-ffi) run exactly the enrollment pipeline the CLI runs — same cleanup stages, same encoder, same finiteness refusal — instead of a parallel one that quietly skips the denoiser.

Errors

When feature extraction or encoding fails, or the encoder yields a non-finite vector.