[BUG] Diarization chunk boundaries split words mid-token in meeting transcripts

[timthinks]

Version: Muesli 0.8.1 (Homebrew cask)

macOS: 26.5.1, Apple Silicon, 24 GB

Transcription model: Parakeet v3 (parakeet-tdt-0.6b-v3-coreml), FluidAudio backend

What happens

When a meeting transcript is diarized, speaker-turn boundaries can fall inside a word, splitting it across two turns and attributing the halves to different speakers.

From an imported 49-minute recording:

[00:00:20] Others: . H
[00:00:23] Speaker 2: ow are you too? Um I'm well.

"How" became "H" + "ow" across two speaker turns. Leading punctuation also gets orphaned onto the front of turns (. Hello. Hello.).

Separately: mid-word spacing inside turns

The same transcripts contain words split by a space without a turn boundary, 35 occurrences in one 49-minute file:

obsess ive     healthi est     account ing     complete ly
servic ing     budget ing      replac ing      includ ing

Both were present with Parakeet v2 and slightly more frequent with v3 (30 → 35 occurrences), so this doesn't look model-specific.

Why it matters

Anything parsing the transcript programmatically (via muesli-cli meetings get, or an LLM summarization pass) sees corrupted tokens. It also inflates wordCount, the same audio reports 9,623 words in Muesli versus 9,092 from a reference transcript of the identical file.

Reproduce

1. Import a multi-speaker audio file via Import Audio (mine was a 49-min mp3, 3 speakers, single mixed track).
2. Read the result: muesli-cli meetings get <id> and inspect rawTranscript.
3. Grep for \b\w+ (?:ive|est|ing|tion|ly)\b and for turns beginning with a partial word.

Happy to share more detail on request, the source recording is confidential so I can't attach it, but I can reproduce against any multi-speaker sample you'd like to nominate.

[timthinks]

Follow-up after testing this on a live recording instead of an import. The behavior is different enough that it may be two separate defects.

Live test: Muesli 0.8.1, Parakeet v2 (parakeet-tdt-0.6b-v2-coreml), 35-minute recorded meeting, 3 people on a Teams call, 5,193 words. Recording saved as mono 16 kHz AAC at roughly 24 kbps.

What is different live

The word splitting still happens, but it takes a different shape and it is confined to one stream. 8 occurrences, all on You (the local microphone). Zero on Speaker 1, Speaker 2, or Others.

H ow      M oney     B bo        F un
w have    f seems    b basically s you

These are single leading characters severed off the front of a word. The imported-audio case in the original report was the opposite shape, a suffix break at a morpheme boundary (account ing, obsess ive, healthi est). Same symptom class, but the split point is not the same, and live it only affects the mic path.

Diarization itself was much better live. The same kind of meeting produced 4 labels on a live 3-person call, versus 8 to 9 labels on an imported single mixed track. So the chunk-boundary problem in the original title looks specific to the import path. The remaining live issue is the mic-stream character splitting above, which may not be a diarization boundary problem at all.

Suggested split if it helps triage:

1. Import path: diarization chunk boundaries cut words and orphan punctuation across turns.
2. Live path: single leading characters severed from words on the local microphone stream only.

Reproduce for 2: record a live meeting with a mic and system audio, then grep rawTranscript for \b[A-Za-z] [a-z] on turns labeled You.

[mvanhorn]

Implemented in #446.