radio: bumper detection in diarizer + full archive download script

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>
This commit is contained in:
2026-04-27 16:17:50 -07:00
parent a4f527f31e
commit c760e430c0
11 changed files with 361 additions and 139 deletions

View File

@@ -15,7 +15,38 @@ The show has had multiple **co-hosts** rotating through, plus **producers / boar
|---|---|---|---|
| **Randall** | early years | not yet | no |
| **Rob** | early years + appearances in 2018/2019 (Mike unsure of exact dates) | not yet | no |
| **Tara** | confirmed 2014-s6e19, 2016-s8e43; diarizer also found her in 2017-s9e30 (610s/11.4%) — pending Mike spot-check | yes | yes — `voice-profiles/tara/` (44 embeddings) |
| **Tony** | 2012-era co-host (Mike unsure whether on-air in 2012-06-09-hr1) | not yet | no |
| **Tara** | confirmed 2014-s6e19, 2016-s8e43, **2018-s10e18 @ 50:50** (verified by Mike 2026-04-27 listen). Plausible in 2015 and 2017 (pending verify). | yes | yes — `voice-profiles/tara/` (44 embeddings, **possibly contaminated**, see below) |
### Tara profile contamination flag
Mike spot-checked CO-HOST-flagged windows on 2026-04-27 and found the diarizer matching:
In **2018-s10e18**:
- **A bumper** (09:20-10:05, music/promo — not a voice)
- **Tara** (50:50 — true positive)
- **A caller, "Christopher"** (~82:10 — false positive, real caller misattributed as Tara)
In **2012-06-09-hr1**:
- **A caller, "Kay"** (22:10-26:00 — real caller misattributed as Tara). Spans the 22:25-24:30 (125s) and 25:15-25:55 (40s) CO-HOST turns. Mike unsure whether co-host Tony was on-air this episode.
In **2015-s7e19** (Jan 2015 New Year episode):
- **A caller, "William"** (~35:30 — confirmed in transcript: "let's talk to William. Hello, William. How are you?", asks about Excel→Word mail merge)
- **A caller, "Charles"** (~16:30 — Mike-identified, transcript not yet verified)
- **A recurring special guest, "Clay" from "Nerd Junkies"** — appears multiple times: transcript at 33:13 "More Clay from the Nerd Junkies", at 37:33 "I'm just curious, Clay, do you have any feedback". Clay is a recurring guest, not a co-host. The 4:40 of "Tara"-attributed audio in this episode is likely **all** Clay + callers, with no actual Tara presence.
### Recurring guests / fill-ins
| Person | Affiliation | Confirmed in audio | Profile built |
|---|---|---|---|
| **Clay** | "Nerd Junkies" — fills in for Tara when she's out (Mike: rarely appears in other episodes) | 2015-s7e19 (throughout — Tara was out, Clay covered) | pending |
Tara's role is explicit per transcript at 2015-s7e19 @ 00:51: "in Tara's place, we have Clay. Clay from the Nerd Junkies." — Tara is the regular co-host for that era; Clay is a fill-in.
Root cause is likely contamination in `build_cohost_profile.py`: the TARA_WINDOWS were sourced from "first 60 min CALLER turns" under the assumption "real callers don't call in during the first hour of a 2-hour show." That assumption appears to leak — at least one real caller ended up in Tara's training data, and the resulting profile now matches a too-broad acoustic space.
Two distinct fixes needed:
1. **Bumper handling in diarizer** — the qa_extractor has bumper signature detection but the diarizer doesn't filter music/promo segments before speaker matching. Bumpers with vocal content can trigger speaker matches.
2. **Tara profile rebuild from vetted windows** — Mike-confirmed windows only, not the heuristic-selected first-60-min approach. The 2026-04-27 listen confirmed 50:50 in 2018-s10e18 as a clean Tara window; more would be needed.
### Producers / board ops (sometimes on-air)
| Person | Profile built |
@@ -23,6 +54,7 @@ The show has had multiple **co-hosts** rotating through, plus **producers / boar
| **Andrew** | no |
| **Shannon** | no |
| **Ken** | no |
| **Unknown board op (2015-s7e19 opening)** | no — Mike heard him at the very start of 2015-s7e19, name forgotten |
| (Mike: "a couple more" he doesn't recall off-hand) | no |
Mike: "The 'producer' (board op) would also be on-air sometimes." Anywhere a producer's voice appears, they're currently being labeled CALLER, which inflates Q&A false positives. Same problem as unprofiled co-hosts.