Merged-decoder (use_cache_branch/past_key_values) loop discovered from the
graph at load; step 1 runs the no-cache branch over the prompt, later steps
feed one token + cached KVs (decoder slots refreshed, encoder cross-attn KVs
kept from step 1). Decoder session lands on the Intel iGPU via the OpenVINO
GPU plugin (same staged runtime) with CPU-EP fallback and a
WA_ONNX_DECODER_DEVICE=cpu|gpu override; DirectML backends keep the CPU
decoder. Legacy decoder_model.onnx is removed once the merged graph loads.
Measured (Core Ultra 5 135U, base.en, 19s TTS clip, identical transcripts):
decoder on OpenVINO/GPU infer=1288ms vs CPU-EP infer=6098ms
- live diarization tick on split recordings now uses attribute_split (mic
channel = 'You' from the first pass) instead of whole-mix clustering +
voiceprint matching, which never reliably surfaced 'You' before stop
- attribute_split assigns via diarization::assign_split (channel-total
overlap) so the user's speech can't be swallowed by a longer far-side span
- stop_recording overlays live user renames over the automatic name map
- rename_speaker/map_speaker_to_participant rewrite notes.md's baked-in
**Name:** dialogue tags so the Notes pane follows the Speakers pane
- model_used records the ONNX model when the NPU/DirectML engine is routed
(was reporting the unused ggml model id) + info log naming the encoder EP
fold_wav folds a WAV's channels to their average across N output channels.
serve_recording plays it as mono so a split (mic-left/loopback-right) recording
plays both sources in both ears; bundle export decrypts + folds to dual-mono so
the shared file is a normal stereo file (manifest records it as summed). Test
covers mono + dual-mono.
stop + reprocess now, for a split recording, diarize the right (loopback)
channel into Speaker N and take You from left-channel (mic) VAD — read back from
the file so both agree. Fixes reprocess collapsing to one speaker. Summed
recordings keep the whole-signal + voiceprint fallback. Removes MicTimelineFile,
phase3_attribute, mask_ranges, mic_activity.json (retirement note retained).
Add a channel-selective 16k reader (share the decode with read_wav_mono_16k via
read_wav_16k) and vad_spans (100ms energy-windowed You detection). Remove the
MicActivity live timeline + its capture wiring — the split layout makes the
mic/far-side separation intrinsic to the file, so You is recomputed from the
left channel instead.
New split writer + a `split` flag through the capture path: when the mic is on,
audio.wav is recorded 2ch with the mic on the left channel and the loopback
(downmixed) on the right, keeping the two sources separate in the file. Same
byte size as today's summed stereo. Transcription still downmixes L+R (= the
summed signal). Unit test covers L=mic / R=loopback.
The capture-silence fix metered the mic pull by wall-clock, but the mic fills
the bridge at its own device clock. Over a long silent stretch they drifted, the
bridge backed up to its 0.5s cap and dropped oldest samples -> the recording
skipped/'fast-forwarded' late in the gap. And a 10ms idle threshold punched
silence into the loopback on brief inter-packet stalls -> audible ticks. Now:
drain the whole bridge (device-paced, no overflow) and only synthesize after a
sustained 150ms gap.
Factor the Phase 3 attribution (you_spans -> mask -> diarize far side -> merge +
name) into a shared phase3_attribute() used by both stop and reprocess. stop now
writes mic_activity.json (the 'You' spans) next to a retained audio.wav
(plaintext timing metadata, local-only, lives/dies with audio.wav); reprocess
prefers that timeline over the voiceprint fallback so re-transcription keeps
correct per-stream 'You'/'Speaker' attribution.
stop_recording now, when the mic is on, takes 'You' spans straight from the
MicActivity timeline and diarizes the recording with those ranges zeroed, so
sherpa only ever clusters the far side into Speaker N. Merges the span lists and
names them via the existing voiceprint map (You + Speaker 2..). Falls back to
the whole-signal pass + voiceprint match when the mic is off. Adds mask_ranges
helper + tests.
Records, per 100ms of audio.wav frame time, when the mic (the user) was
speech-level — captured in the loopback writer where mic and loopback exist
separately in the recording's own timebase. you_spans() collapses active
windows into merged 'You' ranges. The far side will be diarized with these
ranges masked out, so clustering never sees the mic. Struct + span logic land
first (tested); capture wiring + stop attribution follow.
WASAPI loopback delivers no packets when the render endpoint is idle, so the
recording only advanced while system audio played. The bridged mic (the user
talking over silence) was lost to MicBridge overflow, and audio.wav lost those
stretches entirely — corrupting diarization and dropping the user's solo
remarks in real meetings. When loopback is idle, synthesize silence for the
elapsed gap and fold the mic in, keeping audio.wav continuous. Loopback +
mic-enabled only.
VoiceSample kept the first N samples unconditionally, so a user silent for the
first 30s got a voiceprint of room noise/a cough. That noise embedding
spuriously matched a cluster and mislabeled every speaker 'You', even during
silence. Skip sub-threshold (VOICE_ENERGY_FLOOR) chunks so the window tracks
actual speech; until ~1s of real voice accumulates, the match returns empty
(no premature 'You').
Reprocess re-diarizes and upserts the new labels but never deleted the old
ones, so get_meeting (Participants pane) kept returning the discarded run's
labels (e.g. 83) while the transcript showed the new 5. Clear the meeting's
speaker rows before finalize re-inserts the fresh set.
iamdoubz