claudetools/projects/radio-show/session-logs/2026-04-27-qa-extraction-cohost-indexing.md

Session Log: Q&A Extraction — Co-Host Profile + Archive Indexing

Date: 2026-04-27 Project: Radio Show Archive Mining — Computer Guru Show

---

User

• User: Mike Swanson (mike)
• Machine: DESKTOP-0O8A1RL
• Role: admin

---

Session Summary

The session began with resuming work following a benchmark run that demonstrated a significant performance improvement in Whisper transcription, achieving 63.8x real-time speed with batched inference and int8_float16 settings. Next, the focus shifted to evaluating the quality of Q&A extraction across six test episodes, revealing a critical issue with false positives due to co-host Tom being mislabeled as CALLER based on a voice similarity threshold.

A co-host voice profile for Tom was constructed using 44 embeddings from two specific episodes (2014-s6e19 and 2016-s8e43), producing a cosine similarity of 0.698 against Mike — well below Mike's 0.85 threshold, giving clean separation. Code was updated in voice_profiler.py and diarizer.py to correctly emit "Cohost: Tom" labels and map them to a new "CO-HOST" speaker tag. Re-diarizing the two co-host-era episodes dramatically cleaned up Q&A results: 2016 went from 12 false positives to 2 real WiFi caller pairs.

Several bugs in qa_extractor.py were fixed: overlap resolution for sliding-window diarization boundaries, CALLER-preference threshold for long batch transcript segments, and a turn-based caller-intro lookback to replace an ineffective 120s time window. Phone-greeting detection and new promo signatures were added. The final Q&A count landed at 10 pairs across 6 episodes, with 2014 correctly yielding 0 (gaming co-host episode with no actual callers).

archive.db was created with the ArchiveIndex schema (episodes, segments, segments_fts, qa_pairs, qa_fts). All 6 test episodes were indexed: 762 segments, 10 Q&A pairs. FTS5 search verified working for "router", "Windows 10", "Internet Explorer", "antivirus", and "connect" queries.

---

Key Decisions

• Co-host threshold uses same 0.85 bar as host: Tom scores 0.698 vs Mike. Any voice >= 0.85 against Tom's composite gets labeled CO-HOST. Keeps the same single threshold for all profiles rather than per-profile thresholds.
• Turn-based lookback for caller-intro (2 HOST turns, not 120s): Long HOST monologue blocks (8-10 min) in big show segments meant time-based lookback missed the caller introduction. Previous 2 HOST turns always catches it regardless of block length.
• CALLER-preference at 4s minimum overlap: Batch transcription produces ~26s segments; diarization CALLER windows are ~10s. Pure majority-vote always gave HOST. 4s minimum CALLER coverage labels the segment CALLER without being overly aggressive for co-host episodes.
• Midpoint boundary resolution at load time: Rather than re-diarizing everything, the sliding-window overlap is resolved in load_diarized_transcript() so it applies retroactively to all saved diarization files without touching the JSON.
• 751-1041 added as promo signal: Earlier Tucson show number (vs 790-2040 in later seasons). Weighted 1 (needs a second semi-generic signal to filter).
• Tom's windows sourced from first 60 min of co-host episodes: Real callers don't call in during the first hour of a 2-hour show (only exceptions: very end of show). First-hour CALLER windows are safely all Tom.

---

Problems Encountered

• 2016-s8e43 had 12 Q&A pairs, 11 false positives: Root cause was Tom (co-host) labeled CALLER throughout. Fixed by building Tom's voice profile and re-diarizing.
• 2014-s6e19 had 2 Q&A pairs from gaming discussion: Same co-host issue. After re-diarization: 0 pairs (correct — no actual callers in that gaming special).
• 2012-03-10 yielded 0 segments labeled CALLER: Midpoint assignment hit HOST turns (HOST 0-20s and CALLER 15-30s — midpoint 15.1s falls in HOST). Fixed by overlap-preference assignment with 4s CALLER minimum.
• Real WiFi caller (2016, ~4794s) was missing after first fix attempt: Aggressive time-based lookback (120s) combined with short CALLER turns from sliding-window diarization caused the caller question to land in a HOST segment. Fixed by turn-based lookback + co-host profile (eliminated Tom noise, letting real caller windows survive).
• 2012-Jun pair at 1325s was a promo: "The Computer Guru. We'll get your problem solved. Call 751-1041 today" passed promo filter. Fixed by adding 751-1041 and "we'll get your problem solved" as promo signatures.

---

Files Created / Modified

New files

projects/radio-show/audio-processor/build_cohost_profile.py
projects/radio-show/audio-processor/index_test_episodes.py
projects/radio-show/audio-processor/archive.db
projects/radio-show/audio-processor/voice-profiles/tom/
projects/radio-show/audio-processor/voice-profiles/profiles.json  (updated: Tom added)
projects/radio-show/session-logs/2026-04-27-qa-extraction-cohost-indexing.md  (this file)

Modified

src/voice_profiler.py       — emit "Cohost: <name>" label for cohost role
src/diarizer.py             — map "Cohost:" prefix to "CO-HOST" speaker
src/qa_extractor.py         — overlap resolution, CALLER-preference, turn-based
                              caller-intro lookback, _preceded_by_caller_intro(),
                              _PHONE_GREETING, 751-1041 + promo sig additions
test-data/transcripts/2014-s6e19/diarization.json   (re-diarized with Tom profile)
test-data/transcripts/2016-s8e43/diarization.json   (re-diarized with Tom profile)

---

Benchmark Results (from previous run — baseline for BEAST comparison)

Machine: DESKTOP-0O8A1RL — NVIDIA GeForce RTX 5070 Ti Laptop GPU

Episode	Audio	Wall (diarize)	RTF
2011-03-12-hr1	2509s	15.1s	166.1x
2012-03-10-hr1	2634s	12.2s	215.5x
2012-06-09-hr1	2648s	12.2s	216.8x
2014-s6e19	2914s	13.4s	216.9x
2016-s8e43	5326s	24.2s	219.6x
2017-s9e30	5343s	24.7s	216.4x
TOTAL	21374s	101.9s	209.7x

Transcription (batched Whisper large-v3): 63.8x realtime
Diarization: 209.7x realtime
vs DESKTOP-0O8A1RL baseline (149.5x): +60.2x (+40.3%)

---

Archive DB State

Path: projects/radio-show/audio-processor/archive.db

Episodes : 6
Segments : 762
Q&A pairs: 10

Q&A pairs by episode:

Episode	Pairs	Notes
2011-03-12-hr1	3	IE lockout call, cloud computing, ghost hunting caller
2012-03-10-hr1	1	iPad 3 discussion
2012-06-09-hr1	1	Windows repair feature call
2014-s6e19	0	Gaming co-host special — no actual callers
2016-s8e43	2	WiFi connectivity caller (2 turns of same call)
2017-s9e30	3	Software control, Cat5 cabling (Charlie), WiFi ports

---

Voice Profiles State

Path: projects/radio-show/audio-processor/voice-profiles/

Name	Role	Embeddings	Source Episodes
Mike Swanson	host	180	9 episodes (2010-2018)
Tom	cohost	44	2014-s6e19, 2016-s8e43

Tom vs Mike cosine similarity: 0.698 (well-separated at 0.85 threshold)

Tom's source windows used:

• 2014-s6e19: 195-260s, 320-425s, 600-650s, 675-710s
• 2016-s8e43: 100-115s, 135-160s, 270-295s, 575-605s, 1185-1235s, 1790-1870s, 2020-2055s

---

Co-Host Era Notes

Tom was the regular in-studio co-host/board-op roughly 2013-2016. His voice is in episodes from at least 2014 through 2016 (confirmed from test set). The 2011 and 2012 episodes are pure call-in format with no co-host.

If there are occasional guest co-hosts or fill-in hosts in other years, they would still be labeled CALLER until profiled. These would be rare and would likely not form question patterns that survive the caller-intro gate.

---

Pending Tasks for BEAST (GURU-BEAST-ROG)

1. Run benchmark.py to establish RTX 4090 baseline

cd D:/claudetools/projects/radio-show/audio-processor
.venv/Scripts/python benchmark.py 2>&1 | tee bench-4090.txt

BENCH_SETUP.md has all setup steps. The voice profiles are in voice-profiles/ (already copied or available via Tailscale/robocopy from DESKTOP-0O8A1RL). Test episodes go in test-data/episodes/.

Expected: diarization RTF should be ~250-300x on RTX 4090 (vs 209.7x on laptop 5070 Ti). Transcription should be ~70-80x.

Update benchmark.py line 27 after measuring:

BASELINE_RTF  = 209.7  # current laptop 5070 Ti baseline

2. Download full archive from IX server (172.16.3.10)

Use paramiko (SSH with key agent disabled):

import paramiko
ssh = paramiko.SSHClient()
ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
ssh.connect("172.16.3.10", username="gurushow", password="<from vault>",
            look_for_keys=False, allow_agent=False)

Archive path: /home/gurushow/public_html/archive/Radio/ Episode count: 579 MP3s across 2010-2018 (no 2013 season) Approximate total size: ~30-40 GB

Download script skeleton in prior session log: 2026-04-27-diarization-pipeline.md

Tailscale required — IX server is at 172.16.3.10, requires VPN.

3. Full archive processing

Once episodes are downloaded:

# Transcribe + diarize all episodes
cd D:/claudetools/projects/radio-show/audio-processor
.venv/Scripts/python diarize_training.py  # or a new batch_process_all.py

# Index everything into archive.db
.venv/Scripts/python index_test_episodes.py  # modify to point at full episodes dir

The pipeline is idempotent — add_segments() skips episodes already indexed.

4. Verify co-host era episodes

2013-2016 era episodes should now correctly separate Tom (CO-HOST) from actual callers. Spot-check a few 2015 episodes after processing to confirm Tom's profile generalizes well.

If any 2015/2016 episodes show too many CALLER turns that are clearly Tom (voice changed slightly over years), re-run build_cohost_profile.py with windows from that episode added to TOM_WINDOWS dict.

---

Technical Reference

Key thresholds

host_match_threshold = 0.85    # WavLM cosine similarity — applied to ALL profiles
CALLER_MIN_S = 4.0             # min CALLER coverage in transcript segment to label CALLER
PROMO_SCORE_THRESHOLD = 2      # weighted promo signature score
MIN_QUESTION_DURATION = 5.0    # seconds
MIN_ANSWER_DURATION = 15.0     # seconds
MAX_GAP_BETWEEN_QA = 30.0      # seconds

Diarization sliding window

window_s = 10.0   # 10s embedding windows
hop_s = 5.0       # 5s hop → overlapping boundaries (resolved at load time)

Transcription (batch mode)

model_size = "large-v3"
compute_type = "int8_float16"
batch_size = 16
# No word timestamps in batch mode (not needed for search/diarization)

DB search examples

from src.indexer import ArchiveIndex
from pathlib import Path

with ArchiveIndex(Path("archive.db")) as idx:
    # Segment search
    results = idx.search("router", limit=20)
    results = idx.search("Windows 10", speaker_filter="HOST", limit=10)

    # Q&A search
    qa = idx.search_qa("antivirus", limit=10)
    qa = idx.search_qa("wifi connect", limit=10)

Archive server

Host: 172.16.3.10 (requires Tailscale)
User: gurushow
Archive root: /home/gurushow/public_html/archive/Radio/
SSH: paramiko with look_for_keys=False, allow_agent=False