Append to 2026-04-28-session.md covering the FastAPI/SQLite container
deploy: build + ship + verify, plus credentials, paths, and re-deploy
procedures for both DB updates and source updates.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Read-only HTTP layer over archive.db. Endpoints: /api/stats,
/api/episodes, /api/episodes/{id}, /api/episodes/{id}/transcript,
/api/search (FTS5 over segments + qa_pairs, bm25-ranked, snippets),
/api/callers. Single-file HTML index with debounced search UI.
Deployed: Jupiter (Unraid Docker), bound to 172.16.3.20:8765, LAN only.
Container path: /mnt/user/appdata/radio-archive/{app,data}.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Execution-only follow-on to 2026-04-27. Both batch passes done (519+53,
0 errors), import_to_sqlite.py run incrementally to bring archive.db
to final state. Next step: Jupiter Docker container deploy.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
QAPair gets caller_name and caller_role fields populated by a new
attach_caller_names(pairs, transcript_segments) helper. For each pair,
finds the active opening intro at the question_start time (8s forward
tolerance, no backward limit — a caller's call can run for 10+ minutes
and the intro happens once at the start) and attaches the speaker name.
Validation on 9-episode test set:
19/19 Q&A pairs (100%) now have caller names attached.
Examples of corrections from oracle attribution:
2018-s10e18 @ 73:36 Christopher (was misattributed to "Tara")
2015-s7e19 @ 35:45 William (was misattributed to "Tara")
2010-05-08-hr1 Jackie x3, Bruce
2012-03-10-hr1 Adam x2
2016-s8e43 John, Doug
2017-s9e30 Tom, Denise x3, Charlie
speaker_oracle.py: adds speaker_at(time, intros) helper used both by the
existing resolve_speakers() and the new caller-name attachment. Also
adds the "let's fit/bring/put X in/on" intro pattern variant (caught
Charlie at 70:21 in 2017-s9e30 that "talk to X" missed).
download_full_archive.py: SSH keepalive every 30s + per-file retry-on-
failure (up to 3 attempts with reconnect). Earlier run hung on a dead
connection at file 109 of 589 with no recovery; restarted run is now
running at ~10 MB/s vs ~2-3 MB/s before.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New module src/speaker_oracle.py extracts speaker introductions from
transcripts ("let's talk to William", "we have Clay from the Nerd Junkies",
"in Tara's place, we have Clay", "thanks for the call <name>") and binds
them to non-HOST diarization turns. Pure post-pass on diarization JSONs,
no audio processing — corrects audio-only cosine errors using Mike's
deterministic on-air announcements.
Algorithm:
- Extract intros: regex patterns for caller pickups, guest intros,
fill-in announcements, caller closes. Case-strict (rejects mid-sentence
lowercase matches), with a blacklist of common false-positive words.
Deduplicates same-name intros within 5s.
- Resolve speakers: for each non-HOST turn, find the LATEST opening intro
at or before turn.start (with 8s forward tolerance for boundary slop).
Later intros implicitly close earlier callers, so the most recent
intro wins. No artificial lookback limit (callers can talk for 10+ min).
- Falls back to caller_close patterns within 30s after a turn ends.
Validation on 9-episode test set:
2018-s10e18: Christopher 190s correctly named (was mislabeled "Tara")
2012-06-09 : Kay 160s correctly named (was mislabeled "Tara")
2015-s7e19 : Clay 45s as fillin for Tara, William 40s as caller
2016-s8e43 : Charles 630s, Bruce 210s, John 205s — most callers named
2017-s9e30 : Denise 295s, Tom 115s, Elaine 85s, Jeff 10s
Many other callers across all episodes correctly named.
Remaining unnamed CO-HOST/CALLER (~5-10% of non-HOST time) are real
co-host banter or callers without explicit Mike-introductions.
benchmark.py: adds Phase 2.5 "Name Resolution" between diarization and
Q&A extraction. Prints named-speaker breakdown per episode. Doesn't
modify diarization JSONs (resolution is computed on demand).
Next step: feed named turns into qa_extractor so Q&A pairs get caller
name attached for searchability. Also: bootstrap recurring-speaker
profiles (Tara, Tony, Rob, Randall, producers) by accumulating
intro-tagged windows across the full archive once download completes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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>
Added 2010, 2015, 2018 test episodes to round out the test set to one
per available year:
- 2010-05-08-hr1 (May 2010, earliest available; pre-Tara era)
- 2015-s7e19 (Jan 2015, avoids training's s7e30)
- 2018-s10e18 (only 3 non-training 2018 episodes exist)
Archive has no 2019 directory — Rob's "2018/2019 appearances" are
constrained to the 5 available 2018 episodes only.
Per-year diarization summary (Tara presence, post-rename):
2010-05-08 30s 1.2% likely false positive (pre-Tara)
2011-03-12 140s 5.6% likely false positive (call-in only)
2012-03-10 30s 1.1% likely false positive (call-in only)
2012-06-09 340s 12.8% suspicious — Mike to confirm
2014-s6e19 680s 23.3% confirmed
2015-s7e19 280s 9.9% plausible — Mike to confirm
2016-s8e43 1890s 35.5% confirmed
2017-s9e30 610s 11.4% plausible
2018-s10e18 880s 17.1% COULD BE ROB — Mike flagged Rob for
2018/2019 appearances; cosine threshold may
be hitting on Rob being acoustically similar
to Tara
Total Tara across 9 episodes: 1h 21m / 8h 52m audio (15.3%).
Q&A counts (still suspect — every voice that isn't Mike-or-Tara is
labeled CALLER, so Randall/Rob/producers inflate the bucket):
2010=4, 2011=1, 2012a=2, 2012b=0, 2014=0, 2015=1, 2016=2, 2017=4, 2018=3
Total: 17 pairs across 9 episodes
4090 perf on the expanded set:
- Diarization: 31928s in 121.5s = 262.7x realtime (vs 209.7x on 5070 Ti, +25.3%)
- Transcription (3 new episodes only): 10554s in 112.4s = 93.9x
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>
Created Mac M4 batch transcription using mlx-whisper with Apple Silicon
GPU acceleration. Transcribed 8 remaining episodes (17,555 total segments).
Scripts:
- batch_transcribe_mac.py: Full batch processor with mlx-whisper
- test_mac_transcribe.py: Quick test script for faster-whisper
Transcripts (JSON, SRT, TXT formats):
- 2011-06-04-hr1: 1,503 segments
- 2011-09-10-hr1: 1,378 segments
- 2014-s6e05: 1,340 segments
- 2015-s7e30: 1,053 segments
- 2016-s8e42: 2,205 segments
- 2017-s9e26: 2,366 segments
- 2018-s10e17: 4,683 segments
- 2018-s10e21: 2,493 segments
All 9 episodes now transcribed (8 on Mac + 1 from Linux).
Ready for Stages 3-6 on Linux PC.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
GPU firmware bug (NVRM 0x00000062) on RTX 5070 Ti makes
GPU transcription impossible. Handoff doc for Mac M4 to
build native version and complete the 8 remaining episode
transcriptions.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fine-grained speaker analysis (3s windows, 1s hop) across 42min episode
- Host voice: 0.90-0.98 similarity (clear positive match)
- Callers: 0.65-0.68 (correctly below threshold)
- Produced audio/clips: 0.53-0.65 (correctly identified as non-host)
- Co-host/other speakers: 0.56-0.62 (correctly identified)
- Tuned host_match_threshold from 0.75 to 0.83 based on empirical data
- Cross-referenced dips with transcript: correctly identifies callers,
show intros, played audio clips, and station breaks
- Batch transcription of 7 additional training episodes in progress
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Voice profiler using microsoft/wavlm-base-sv (512-dim x-vector embeddings)
- Bootstrap from archive: 180 embeddings from 9 episodes across 2010-2018
- Host identification accuracy: 0.87-0.98 similarity for live speech,
0.60-0.64 for non-host audio (produced intros, co-host)
- Dropped speechbrain dependency (requires torchaudio, CUDA version conflicts)
- Patched torchaudio CUDA 12.8/13.1 version check (warning instead of error)
- Profile stored in voice-profiles/mike-swanson/ with per-chunk embeddings
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Audio processor CLI tool with 6-stage pipeline: transcribe (faster-whisper GPU),
diarize (pyannote), detect segments (multi-signal classifier), remove commercials,
split segments, analyze content (Ollama)
- Post-show workflow doc for episode posts, forum threads, deep-dive blog posts
- Training plan for using 579-episode archive for voice profiles and commercial detection
- Successful test: 45min episode transcribed in 2:37 on RTX 5070 Ti
- Sample transcript output from S7E30 (March 2015)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
