fix: normalize i16 to f32 in segmentation and flush trailing speech

[gregoire22enpc]

Problem

get_segments in segment.rs has two bugs:

1. Missing normalization in i16→f32 conversion (line 59)

// Current (broken):
let array = ndarray::Array1::from_iter(window.iter().map(|&x| x as f32));

// Fixed:
let array = ndarray::Array1::from_iter(window.iter().map(|&x| x as f32 / 32768.0));

The ONNX segmentation model was trained on input in [-1.0, 1.0], but x as f32 feeds it raw i16 values in [-32768, 32767]. The model's weights, biases, and normalization layers are calibrated for the expected range, so values 32768x larger produce unreliable frame classifications — in practice, all frames are classified as non-speech.

This was discovered while integrating pyannote-rs in silverstein/minutes, where minutes enroll consistently failed with "No speech detected in the recording" and minutes record produced transcripts with missing speaker labels.

2. Trailing speech segment dropped at end-of-audio

When speech extends to the end of the audio (i.e., is_speeching is still true after all windows are processed), the final segment is silently dropped because there is no flush at the end of the iterator. This fix emits the trailing segment.

Changes

• segment.rs line 59: Divide by 32768.0 when converting i16 to f32
• segment.rs end of iterator: Flush the last speech segment when is_speeching is still true after processing all windows

Verification

cargo check passes. The fix is a minimal, targeted change to the existing logic.

[gregoire22enpc]

@thewh1teagle, happy to get your thoughts on this :)

[silverstein]

Hi @thewh1teagle, we're using pyannote-rs in Minutes and this fix unblocks voice enrollment for our users. We're currently pinned to a fork with this patch. Happy to help test or co-maintain if that would be useful. Any timeline for a release?