First attempt at Clay's voice profile from 2015-s7e19 produced
Clay-vs-Mike cosine similarity of 0.994 — essentially a Mike clone.
Root cause: 10s WavLM x-vector chunks averaged Mike's frequent
interjections together with Clay's dialogue, and Mike's well-trained
profile dominated the resulting embedding signal.
Mike's call: skip Clay, accept the 2015-s7e19 Q&A as noisy. Clay rarely
appears in other episodes, so the cost of not having his profile is
bounded to this one episode plus any rare future appearances.
Cleanup:
- voice-profiles/clay/ removed
- voice-profiles/profiles.json: Clay entry removed
- Memory updated to record the decision and the failure mode
Kept build_clay_profile.py in-repo as documentation of the attempt and
the Mike-similarity-filter pattern. Useful starting point if a future
attempt provides cleaner pure-Clay timestamps.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a transcript-driven bumper filter to the diarization pipeline. When
a transcript segment matches qa_extractor's promo/bumper signatures, the
overlapping audio windows are labeled BUMPER and the WavLM cosine match
is skipped. Prevents music/promo from being matched against speaker
profiles (the failure mode Mike caught in 2018-s10e18 @ 09:20-10:05).
Code changes:
- src/voice_profiler.py: identify_speakers() takes optional skip_ranges
parameter; windows whose midpoint falls in a skip range get labeled
"[bumper]" and skip cosine match
- src/diarizer.py: diarize() takes optional transcript_path; pre-computes
bumper time ranges via qa_extractor._is_promo_or_bumper, passes to
identify_speakers; adds BUMPER speaker label
- benchmark.py: passes transcript_path to diarize()
Aggregate impact across 9-episode test set:
Tara attribution: 4880s -> 3680s (-1200s / -25%)
Q&A pairs: 17 -> 19 (+2)
(bumper-flagged segments had been disrupting conversation detection
in 2017-s9e30 and 2018-s10e18)
CALLER total: 1320s -> 1190s (bumpers previously labeled CALLER moved)
Per-episode bumpers caught: 1-8, total ~165 bumper segments across set
Remaining Tara false positives are real callers acoustically similar to
Tara (Christopher in 2018, Kay in 2012, William and Charles in 2015) and
guest Clay in 2015-s7e19 — those need profile rebuild + Clay profile,
not bumper filtering.
Adds download_full_archive.py — resumable mirror-style downloader that
walks IX server's /home/gurushow/public_html/archive/{year}/ and copies
all MP3s to archive-data/episodes/. Run is in progress (~589 files,
~10-15GB). Used to source clean profile windows for the remaining
co-hosts (Tara rebuild, Clay, Tony, Rob, Randall, producers).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mike confirmed there is no co-host named "Tom" — the voice in 2014-s6e19
and 2016-s8e43 is Tara. The 5070 Ti session fabricated the Tom identity.
The voice profile itself (44 embeddings, 0.698 cosine vs Mike) is correct;
only the human label was wrong.
Rename swept:
- voice-profiles/tom/ -> voice-profiles/tara/ (git mv preserves all .npy)
- voice-profiles/profiles.json: "Tom" key -> "Tara"
- build_cohost_profile.py: TOM_WINDOWS -> TARA_WINDOWS, COHOST_NAME, comments
- 2026-04-27-qa-extraction-cohost-indexing.md: correction header + body sweep
- 2026-04-27-4090-benchmark-and-test-set.md: closure note
- .claude/memory/radio_show_no_cohost_named_tom.md: resolution + speaker roster
Diarization re-run after rename so speaker_map emits "Cohost: Tara".
Q&A counts unchanged (rename is label-only): 9 pairs across 6 test episodes.
Tara distribution from the post-rename diarization (per-episode % of audio):
2011-03-12-hr1 140s 5.6% likely false positive (call-in only)
2012-03-10-hr1 30s 1.1% likely false positive (call-in only)
2012-06-09-hr1 340s 12.8% suspicious — pending Mike confirm
2014-s6e19 680s 23.3% confirmed
2016-s8e43 1890s 35.5% confirmed
2017-s9e30 610s 11.4% plausible — pending Mike confirm
Broader speaker-roster context Mike provided this session (saved to
memory): the show has had multiple co-hosts (Tara, Randall, Rob) plus
producers/board ops (Andrew, Shannon, Ken, others) who would sometimes
go on-air. Only Tara has a profile so far. Every other speaker is
currently labeled CALLER, which means small CO-HOST attributions in
unexpected episodes (e.g. 2011/2012) may actually be a producer rather
than a false positive — Mike to spot-check.
Action item before full-archive run: build profiles for Randall, Rob,
and the named producers to avoid systematic Q&A false positives in
early-years and 2018/2019 episodes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Re-ran benchmark.py on GURU-BEAST-ROG against the post-overhaul code
(co-host profile, batched Whisper int8_float16, revised Q&A extractor).
Results vs 5070 Ti baseline:
- Diarization: 209.7x -> 338.1x (+61.2%)
- Transcription: 63.8x -> 94.8x (+48.6%)
- Q&A pairs: 9 vs 10 (within run-to-run noise; structural correctness matches:
2014 = 0 callers, 2016 = 2 WiFi caller pairs)
Setup change: BENCH_SETUP.md now lists ffmpeg as a Step-2 prereq
(winget install Gyan.FFmpeg). Was missing on this machine and the pipeline
fails silently at the first diarize call without ffprobe.
Code change: benchmark.py BASELINE_RTF updated 149.5 -> 209.7 to reflect
the 5070 Ti's post-overhaul measurement (e9ac607).
Data: 6 test episode transcripts and diarizations regenerated under the
new code path (batched Whisper output + co-host-aware speaker_map).
Correction memory: voice-profiles/tom/ directory + 5070 Ti session log
fabricated a co-host named "Tom" — Mike confirms no such person exists on
the show. The audio profile is real and the diarization separation is
sound, but the human identity attached to it is wrong. Saved under
.claude/memory/radio_show_no_cohost_named_tom.md pending Mike providing
the correct name for rename.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
