azcomputerguru/claudetools
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

radio: bumper detection in diarizer + full archive download script

Browse Source 
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:
Mike Swanson
2026-04-27 16:17:50 -07:00
parent a4f527f31e
commit c760e430c0
11 changed files with 361 additions and 139 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
projects/radio-show/audio-processor/benchmark.py
View File

@@ -108,7 +108,8 @@ for ep, transcript_path, audio_dur, _ in trans_results:
audio_dur = json.load(f).get("duration", 0)
t0 = time.monotonic()
result = diarize(ep, voice_profiles=voice_profiles, host_match_threshold=0.85)
result = diarize(ep, voice_profiles=voice_profiles, host_match_threshold=0.85,
transcript_path=transcript_path)
wall = time.monotonic() - t0
rtf = audio_dur / wall if wall > 0 else 0

103
projects/radio-show/audio-processor/download_full_archive.py Normal file
View File

@@ -0,0 +1,103 @@
"""
Download the full Computer Guru Show archive from IX server (172.16.3.10).
Mirrors the year-based directory structure as-is to archive-data/episodes/.
Resumable: skips files already present with matching size.
Requires Tailscale.
"""
import os
import sys
import time
import paramiko
from pathlib import Path
password = os.environ.get("IX_PASSWORD")
if not password:
print("IX_PASSWORD env var not set", file=sys.stderr)
sys.exit(1)
LOCAL_ROOT = Path(__file__).parent / "archive-data" / "episodes"
LOCAL_ROOT.mkdir(parents=True, exist_ok=True)
REMOTE_ROOT = "/home/gurushow/public_html/archive"
YEARS = ["2010", "2011", "2012", "2014", "2015", "2016", "2017", "2018"]
print(f"Connecting to 172.16.3.10...", flush=True)
client = paramiko.SSHClient()
client.set_missing_host_key_policy(paramiko.AutoAddPolicy())
client.connect("172.16.3.10", username="root", password=password,
look_for_keys=False, allow_agent=False, timeout=30)
sftp = client.open_sftp()
print("Connected.", flush=True)
def list_remote_mp3s(year: str) -> list[str]:
cmd = f"find '{REMOTE_ROOT}/{year}' -iname '*.mp3' 2>/dev/null"
stdin, stdout, stderr = client.exec_command(cmd)
return [line.strip() for line in stdout.read().decode().splitlines() if line.strip()]
total_files = 0
total_bytes = 0
skipped_files = 0
skipped_bytes = 0
downloaded_files = 0
downloaded_bytes = 0
errors = []
t_start = time.monotonic()
for year in YEARS:
print(f"\n=== {year} ===", flush=True)
remote_paths = list_remote_mp3s(year)
print(f" {len(remote_paths)} MP3 files found on remote", flush=True)
for remote in remote_paths:
rel = remote[len(REMOTE_ROOT) + 1:]
local = LOCAL_ROOT / rel
local.parent.mkdir(parents=True, exist_ok=True)
try:
remote_stat = sftp.stat(remote)
remote_size = remote_stat.st_size
except Exception as e:
errors.append(f"stat {remote}: {e}")
continue
total_files += 1
total_bytes += remote_size
if local.exists() and local.stat().st_size == remote_size:
skipped_files += 1
skipped_bytes += remote_size
continue
size_mb = remote_size / 1024 / 1024
print(f" [{downloaded_files + 1:3d}] {rel} ({size_mb:.1f} MB)...", end="", flush=True)
t0 = time.monotonic()
try:
sftp.get(remote, str(local))
elapsed = time.monotonic() - t0
mbps = size_mb / elapsed if elapsed > 0 else 0
print(f" done ({elapsed:.1f}s, {mbps:.1f} MB/s)", flush=True)
downloaded_files += 1
downloaded_bytes += remote_size
except Exception as e:
print(f" FAILED: {e}", flush=True)
errors.append(f"get {remote}: {e}")
elapsed_total = time.monotonic() - t_start
print(f"\n=== Summary ===", flush=True)
print(f" Total remote files : {total_files}", flush=True)
print(f" Total remote bytes : {total_bytes / 1024 / 1024 / 1024:.2f} GB", flush=True)
print(f" Already present : {skipped_files} files / {skipped_bytes / 1024 / 1024 / 1024:.2f} GB", flush=True)
print(f" Newly downloaded : {downloaded_files} files / {downloaded_bytes / 1024 / 1024 / 1024:.2f} GB", flush=True)
print(f" Errors : {len(errors)}", flush=True)
print(f" Wall time : {elapsed_total:.1f}s", flush=True)
if errors:
print(f"\n=== Errors ===", flush=True)
for e in errors[:20]:
print(f" {e}", flush=True)
sftp.close()
client.close()

27
projects/radio-show/audio-processor/src/diarizer.py
View File

@@ -158,12 +158,17 @@ def diarize(audio_path: str | Path,
voice_profiles: VoiceProfileStore | None = None,
min_speakers: int = 1,
max_speakers: int = 6,
host_match_threshold: float = 0.85) -> DiarizationResult:
host_match_threshold: float = 0.85,
transcript_path: str | Path | None = None) -> DiarizationResult:
"""Run speaker diarization using WavLM sliding-window speaker identification.
Uses the built-in VoiceProfiler (WavLM x-vectors) — no HuggingFace token
or gated model required. Identifies HOST vs non-HOST speakers using the
stored voice profile for Mike Swanson.
If transcript_path is provided, time ranges containing show promo/bumper
text are pre-marked and skipped at speaker-identification time so vocal
music doesn't match cohost profiles.
"""
import torch
from .voice_profiler import VoiceProfiler
@@ -190,10 +195,28 @@ def diarize(audio_path: str | Path,
speaker_map={"HOST": "HOST"},
)
# Pre-compute bumper / promo time ranges from transcript if available
bumper_ranges: list[tuple[float, float]] = []
if transcript_path is not None:
transcript_path = Path(transcript_path)
if transcript_path.exists():
from .qa_extractor import _is_promo_or_bumper
with open(transcript_path) as f:
tdata = json.load(f)
for seg in tdata.get("segments", []):
if _is_promo_or_bumper(seg.get("text", "")):
bumper_ranges.append((seg["start"], seg["end"]))
if bumper_ranges:
console.print(
f"[dim]Bumper filter: {len(bumper_ranges)} promo/bumper "
f"transcript segments will be skipped during speaker match[/dim]"
)
# Sliding-window identification: 10s windows, 5s hop
voice_segs = profiler.identify_speakers(
audio_path, window_s=10.0, hop_s=5.0,
threshold=host_match_threshold,
skip_ranges=bumper_ranges,
)
# Convert VoiceSegment labels to HOST / CALLER
@@ -204,6 +227,8 @@ def diarize(audio_path: str | Path,
speaker = "HOST"
elif label.startswith("Cohost:"):
speaker = "CO-HOST"
elif label == "[bumper]":
speaker = "BUMPER"
elif label == "[error]":
speaker = "UNKNOWN"
else:

21
projects/radio-show/audio-processor/src/voice_profiler.py
View File

@@ -279,12 +279,19 @@ class VoiceProfiler:
def identify_speakers(self, audio_path: Path,
window_s: float = 10.0,
hop_s: float = 5.0,
threshold: float = 0.70) -> list[VoiceSegment]:
threshold: float = 0.70,
skip_ranges: list[tuple[float, float]] | None = None
) -> list[VoiceSegment]:
"""Identify speakers throughout an audio file using sliding window.
Loads the full audio once then slices in memory — avoids spawning
hundreds of ffmpeg subprocesses.
Returns timestamped segments with speaker labels and embeddings.
skip_ranges: list of (start, end) seconds. Windows whose midpoint
falls inside any of these ranges are labeled "[bumper]" and the
speaker cosine match is skipped — used to suppress music/promo
from being matched against speaker profiles.
"""
console.print(f"[bold]Identifying speakers:[/bold] {audio_path.name}")
@@ -293,6 +300,8 @@ class VoiceProfiler:
audio = self._load_full_audio(audio_path) # float32 mono array
self._get_model() # ensure model is warm before the loop
skip_ranges = skip_ranges or []
segments = []
window_samples = int(window_s * SAMPLE_RATE)
hop_samples = int(hop_s * SAMPLE_RATE)
@@ -306,6 +315,16 @@ class VoiceProfiler:
s = int(start * SAMPLE_RATE)
e = min(s + window_samples, total_samples)
mid = (start + end) / 2
in_bumper = any(rs <= mid <= re for rs, re in skip_ranges)
if in_bumper:
segments.append(VoiceSegment(
start=start, end=end,
speaker_label="[bumper] (1.00)",
))
continue
try:
emb = self._embed_audio_np(audio[s:e])

21
projects/radio-show/audio-processor/test-data/transcripts/2011-03-12-hr1/diarization.json
View File

@@ -1,26 +1,21 @@
{
"num_speakers": 3,
"num_speakers": 4,
"speaker_map": {
"CALLER": "CALLER",
"HOST": "HOST",
"CO-HOST": "CO-HOST"
"CO-HOST": "CO-HOST",
"BUMPER": "BUMPER",
"CALLER": "CALLER"
},
"turns": [
{
"speaker": "HOST",
"speaker": "BUMPER",
"start": 0.0,
"end": 20.0,
"confidence": 0.89
},
{
"speaker": "CO-HOST",
"start": 15.0,
"end": 25.0,
"confidence": 0.87
"end": 35.0,
"confidence": 1.0
},
{
"speaker": "HOST",
"start": 20.0,
"start": 30.0,
"end": 40.0,
"confidence": 0.88
},

29
projects/radio-show/audio-processor/test-data/transcripts/2012-03-10-hr1/diarization.json
View File

@@ -1,34 +1,23 @@
{
"num_speakers": 3,
"num_speakers": 4,
"speaker_map": {
"CALLER": "CALLER",
"HOST": "HOST",
"CO-HOST": "CO-HOST"
"CO-HOST": "CO-HOST",
"BUMPER": "BUMPER",
"CALLER": "CALLER"
},
"turns": [
{
"speaker": "HOST",
"speaker": "BUMPER",
"start": 0.0,
"end": 20.0,
"confidence": 0.88
},
{
"speaker": "CO-HOST",
"start": 15.0,
"end": 25.0,
"confidence": 0.87
},
{
"speaker": "CALLER",
"start": 20.0,
"end": 30.0,
"confidence": 0.84
"end": 35.0,
"confidence": 1.0
},
{
"speaker": "HOST",
"start": 25.0,
"start": 30.0,
"end": 430.0,
"confidence": 0.86
"confidence": 0.96
},
{
"speaker": "CALLER",

35
projects/radio-show/audio-processor/test-data/transcripts/2012-06-09-hr1/diarization.json
View File

@@ -1,28 +1,23 @@
{
"num_speakers": 3,
"num_speakers": 4,
"speaker_map": {
"CALLER": "CALLER",
"HOST": "HOST",
"CO-HOST": "CO-HOST"
"CO-HOST": "CO-HOST",
"BUMPER": "BUMPER",
"CALLER": "CALLER"
},
"turns": [
{
"speaker": "HOST",
"speaker": "BUMPER",
"start": 0.0,
"end": 20.0,
"confidence": 0.9
},
{
"speaker": "CO-HOST",
"start": 15.0,
"end": 25.0,
"confidence": 0.87
"end": 35.0,
"confidence": 1.0
},
{
"speaker": "HOST",
"start": 20.0,
"start": 30.0,
"end": 690.0,
"confidence": 0.86
"confidence": 0.97
},
{
"speaker": "CALLER",
@@ -33,14 +28,20 @@
{
"speaker": "HOST",
"start": 690.0,
"end": 1350.0,
"end": 1330.0,
"confidence": 0.92
},
{
"speaker": "BUMPER",
"start": 1325.0,
"end": 1355.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 1345.0,
"start": 1350.0,
"end": 1470.0,
"confidence": 0.92
"confidence": 0.93
},
{
"speaker": "HOST",

49
projects/radio-show/audio-processor/test-data/transcripts/2014-s6e19/diarization.json
View File

@@ -1,16 +1,23 @@
{
"num_speakers": 3,
"num_speakers": 4,
"speaker_map": {
"CALLER": "CALLER",
"HOST": "HOST",
"CALLER": "CALLER",
"BUMPER": "BUMPER",
"CO-HOST": "CO-HOST"
},
"turns": [
{
"speaker": "CO-HOST",
"speaker": "BUMPER",
"start": 0.0,
"end": 35.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 30.0,
"end": 40.0,
"confidence": 0.96
"confidence": 0.93
},
{
"speaker": "HOST",
@@ -61,22 +68,28 @@
"confidence": 0.96
},
{
"speaker": "HOST",
"speaker": "BUMPER",
"start": 660.0,
"end": 680.0,
"confidence": 0.98
"end": 695.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 675.0,
"end": 710.0,
"confidence": 0.94
"start": 690.0,
"end": 700.0,
"confidence": 0.95
},
{
"speaker": "BUMPER",
"start": 695.0,
"end": 740.0,
"confidence": 1.0
},
{
"speaker": "HOST",
"start": 705.0,
"start": 735.0,
"end": 985.0,
"confidence": 0.9
"confidence": 0.87
},
{
"speaker": "CO-HOST",
@@ -159,18 +172,18 @@
{
"speaker": "HOST",
"start": 2055.0,
"end": 2155.0,
"end": 2120.0,
"confidence": 0.94
},
{
"speaker": "CALLER",
"start": 2150.0,
"end": 2160.0,
"confidence": 0.83
"speaker": "BUMPER",
"start": 2115.0,
"end": 2165.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 2155.0,
"start": 2160.0,
"end": 2170.0,
"confidence": 0.97
},

81
projects/radio-show/audio-processor/test-data/transcripts/2016-s8e43/diarization.json
View File

@@ -1,14 +1,21 @@
{
"num_speakers": 3,
"num_speakers": 4,
"speaker_map": {
"CALLER": "CALLER",
"HOST": "HOST",
"CALLER": "CALLER",
"BUMPER": "BUMPER",
"CO-HOST": "CO-HOST"
},
"turns": [
{
"speaker": "CO-HOST",
"speaker": "BUMPER",
"start": 0.0,
"end": 35.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 30.0,
"end": 40.0,
"confidence": 0.96
},
@@ -123,14 +130,20 @@
{
"speaker": "HOST",
"start": 550.0,
"end": 580.0,
"end": 565.0,
"confidence": 0.98
},
{
"speaker": "BUMPER",
"start": 560.0,
"end": 595.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 575.0,
"start": 590.0,
"end": 600.0,
"confidence": 0.96
"confidence": 0.93
},
{
"speaker": "CALLER",
@@ -153,12 +166,18 @@
{
"speaker": "HOST",
"start": 1055.0,
"end": 1190.0,
"end": 1160.0,
"confidence": 0.99
},
{
"speaker": "BUMPER",
"start": 1155.0,
"end": 1205.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 1185.0,
"start": 1200.0,
"end": 1215.0,
"confidence": 0.98
},
@@ -255,12 +274,18 @@
{
"speaker": "CO-HOST",
"start": 2020.0,
"end": 2055.0,
"end": 2030.0,
"confidence": 0.92
},
{
"speaker": "BUMPER",
"start": 2025.0,
"end": 2060.0,
"confidence": 1.0
},
{
"speaker": "HOST",
"start": 2050.0,
"start": 2055.0,
"end": 2105.0,
"confidence": 0.98
},
@@ -549,14 +574,14 @@
{
"speaker": "HOST",
"start": 3370.0,
"end": 3395.0,
"end": 3390.0,
"confidence": 0.94
},
{
"speaker": "CO-HOST",
"start": 3390.0,
"speaker": "BUMPER",
"start": 3385.0,
"end": 3435.0,
"confidence": 0.85
"confidence": 1.0
},
{
"speaker": "HOST",
@@ -565,22 +590,16 @@
"confidence": 0.98
},
{
"speaker": "CO-HOST",
"speaker": "BUMPER",
"start": 3965.0,
"end": 3980.0,
"confidence": 0.96
},
{
"speaker": "HOST",
"start": 3975.0,
"end": 3990.0,
"confidence": 0.97
"end": 4020.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 3985.0,
"start": 4015.0,
"end": 4025.0,
"confidence": 0.86
"confidence": 0.95
},
{
"speaker": "CALLER",
@@ -723,14 +742,20 @@
{
"speaker": "HOST",
"start": 4575.0,
"end": 4680.0,
"end": 4655.0,
"confidence": 0.97
},
{
"speaker": "BUMPER",
"start": 4650.0,
"end": 4695.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 4675.0,
"start": 4690.0,
"end": 4715.0,
"confidence": 0.92
"confidence": 0.94
},
{
"speaker": "HOST",

97
projects/radio-show/audio-processor/test-data/transcripts/2017-s9e30/diarization.json
View File

@@ -1,22 +1,17 @@
{
"num_speakers": 3,
"num_speakers": 4,
"speaker_map": {
"CALLER": "CALLER",
"HOST": "HOST",
"CO-HOST": "CO-HOST"
"CO-HOST": "CO-HOST",
"BUMPER": "BUMPER",
"CALLER": "CALLER"
},
"turns": [
{
"speaker": "HOST",
"speaker": "BUMPER",
"start": 0.0,
"end": 20.0,
"confidence": 0.88
},
{
"speaker": "CO-HOST",
"start": 15.0,
"end": 25.0,
"confidence": 0.92
"confidence": 1.0
},
{
"speaker": "HOST",
@@ -69,14 +64,20 @@
{
"speaker": "HOST",
"start": 615.0,
"end": 730.0,
"end": 710.0,
"confidence": 0.89
},
{
"speaker": "BUMPER",
"start": 705.0,
"end": 750.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 725.0,
"start": 745.0,
"end": 770.0,
"confidence": 0.91
"confidence": 0.96
},
{
"speaker": "HOST",
@@ -117,9 +118,21 @@
{
"speaker": "CO-HOST",
"start": 1310.0,
"end": 1355.0,
"end": 1320.0,
"confidence": 0.98
},
{
"speaker": "BUMPER",
"start": 1315.0,
"end": 1350.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 1345.0,
"end": 1355.0,
"confidence": 0.97
},
{
"speaker": "HOST",
"start": 1350.0,
@@ -189,20 +202,20 @@
{
"speaker": "HOST",
"start": 1460.0,
"end": 2130.0,
"end": 2110.0,
"confidence": 0.88
},
{
"speaker": "CALLER",
"start": 2125.0,
"end": 2135.0,
"confidence": 0.78
"speaker": "BUMPER",
"start": 2105.0,
"end": 2155.0,
"confidence": 1.0
},
{
"speaker": "CO-HOST",
"start": 2130.0,
"start": 2150.0,
"end": 2175.0,
"confidence": 0.86
"confidence": 0.89
},
{
"speaker": "HOST",
@@ -219,20 +232,20 @@
{
"speaker": "HOST",
"start": 2650.0,
"end": 2725.0,
"end": 2715.0,
"confidence": 0.97
},
{
"speaker": "CO-HOST",
"start": 2720.0,
"end": 2730.0,
"confidence": 0.89
"speaker": "BUMPER",
"start": 2710.0,
"end": 2745.0,
"confidence": 1.0
},
{
"speaker": "HOST",
"start": 2725.0,
"start": 2740.0,
"end": 2995.0,
"confidence": 0.91
"confidence": 0.99
},
{
"speaker": "CO-HOST",
@@ -273,20 +286,20 @@
{
"speaker": "CO-HOST",
"start": 3375.0,
"end": 3410.0,
"end": 3390.0,
"confidence": 0.91
},
{
"speaker": "CALLER",
"start": 3405.0,
"end": 3415.0,
"confidence": 0.84
"speaker": "BUMPER",
"start": 3385.0,
"end": 3425.0,
"confidence": 1.0
},
{
"speaker": "HOST",
"start": 3410.0,
"start": 3420.0,
"end": 4185.0,
"confidence": 0.96
"confidence": 0.98
},
{
"speaker": "CALLER",
@@ -387,14 +400,20 @@
{
"speaker": "CO-HOST",
"start": 4550.0,
"end": 4595.0,
"end": 4565.0,
"confidence": 0.89
},
{
"speaker": "BUMPER",
"start": 4560.0,
"end": 4605.0,
"confidence": 1.0
},
{
"speaker": "HOST",
"start": 4590.0,
"start": 4600.0,
"end": 5285.0,
"confidence": 0.95
"confidence": 0.94
},
{
"speaker": "CO-HOST",