fix(asr): revert default Whisper model large-v3 → large-v2 (regression)

During v1.8.13 acceptance testing (F4/F6/F7 vs F8 side-by-side on the
293-second JAV reference clip), the v1.8.12 aggressive ASR preset retune
was found to interact pathologically with large-v3 on continuous-energy
non-phonetic content (JAV moaning), producing 6–10 of 68 GT entries
under simple Transcription Mode (~85–90% loss).

The same audio + same preset values + large-v2 produced 51 entries
(75% capture). Diagnostic suite (G_PROD_CL30 variant, production
aggressive params verbatim) confirmed: 33 segments / 12 of 21 groups
with large-v2; equivalent runs with large-v3 catastrophic.

Empirical evidence summary:
  Run                                   Model     Entries / 68 GT
  F4 (whisperseg + auditok + simple)    large-v3        10  ✗
  F6 / F7CLI-GUI (silero + auditok)     large-v3         6  ✗
  F7 mine (silero + auditok)            large-v3         6  ✗
  F5 ensemble pass1 (silero + auditok)  large-v2        52  ✓
  F8 (this fix; silero + auditok)       large-v2        51  ✓

Root cause: the v1.8.12 aggressive preset (no_speech_threshold=0.84,
beam_size=3, best_of=2, temperature=[0.0, 0.17],
compression_ratio_threshold=2.6, repetition_penalty=1.3,
no_repeat_ngram_size=3, chunk_length=30) was tuned against large-v2
forensic acceptance data. large-v3's slightly different encoder/decoder
behavior makes the same gate values too strict on this content
distribution.

Files changed:
  whisperjav/config/components/asr/faster_whisper.py:211
    model_id "large-v3" → "large-v2" (BalancedPipeline default)

  whisperjav/config/components/asr/stable_ts.py:293
    model_id "large-v3" → "large-v2" (Fast/Faster pipeline default)
    Same v1.8.12 retune mindset applies — these presets share lineage
    with the FasterWhisper backend. Aligning defaults avoids
    per-backend divergence.

  whisperjav/webview_gui/assets/index.html:299
    GUI model dropdown selected option "large-v3" → "large-v2"
    Reordered options so large-v2 appears first.

OpenAI-Whisper backend (config/components/asr/openai_whisper.py:178)
was already on large-v2 in v1.8.12 and unchanged.

Users wanting large-v3 can opt in:
  CLI:  --model large-v3
  GUI:  model override checkbox + dropdown selection

v1.9.x will re-run forensic acceptance against large-v3 and produce
a per-engine, per-model preset variant so large-v3 can return as the
default.

Test artifacts: test_media/1813 acceptance/F4 (catastrophic baseline),
F8 (verified fix), F4/DIAG_FW/chunk_length_test (G_PROD_CL30 vs
H_PROD_NOCL diagnostic).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: meizhong986

whisperjav/config/components/asr/faster_whisper.py

@@ -208,7 +208,19 @@ class FasterWhisperASR(ASRComponent):
 
     # === ASR-specific ===
     provider = "faster_whisper"
-    model_id = "large-v3"
+    # v1.8.13: reverted large-v3 → large-v2. Empirical regression discovered
+    # during F4/F6/F7 acceptance testing on dev_v1.8.13: the v1.8.12 aggressive
+    # ASR preset (no_speech=0.84, beam=3, best_of=2, temperature=[0.0, 0.17],
+    # compression=2.6, repetition_penalty=1.3, no_repeat_ngram_size=3,
+    # chunk_length=30) produces catastrophic empty output on JAV content with
+    # large-v3 (6–10 of 68 GT entries) but works correctly with large-v2
+    # (33+ segments captured in diagnostic suite, 52 entries in F5 ensemble
+    # pass1 on the same audio). The v1.8.12 preset retune was tuned against
+    # large-v2 forensic acceptance data; the same values interact pathologically
+    # with large-v3 on continuous-energy non-phonetic content (JAV moaning).
+    # Until the preset is re-tuned for large-v3 in v1.9.x, large-v2 stays the
+    # safer default. Users wanting v3 can opt in with --model large-v3.
+    model_id = "large-v2"
     supported_tasks = ["transcribe", "translate"]
     compatible_vad = ["silero", "faster_whisper_vad", "none"]
 

whisperjav/config/components/asr/stable_ts.py

@@ -290,7 +290,13 @@ class StableTSASR(ASRComponent):
 
     # === ASR-specific ===
     provider = "stable_ts"
-    model_id = "large-v3"
+    # v1.8.13: reverted large-v3 → large-v2 for consistency with FasterWhisper
+    # backend (whose v1.8.12 aggressive preset retune was tuned against large-v2
+    # forensic acceptance data and interacts pathologically with large-v3 on
+    # JAV content — see faster_whisper.py:211 comment for full context). Stable-TS
+    # presets share the same lineage. Aligning the default avoids per-backend
+    # divergence; users wanting v3 can opt in with --model large-v3.
+    model_id = "large-v2"
     supported_tasks = ["transcribe", "translate"]
     compatible_vad = ["none"]  # Stable-TS has built-in VAD
 

whisperjav/webview_gui/assets/index.html

@@ -296,8 +296,8 @@ <h2>Destination</h2>
                                 </div>
                                 <div class="form-group tight-pair">
                                     <select id="modelSelection" class="form-select" disabled>
-                                        <option value="large-v3" selected>large-v3</option>
-                                        <option value="large-v2">large-v2</option>
+                                        <option value="large-v2" selected>large-v2</option>
+                                        <option value="large-v3">large-v3</option>
                                         <option value="turbo">turbo</option>
                                     </select>
                                 </div>