You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
- core.py: set project.subtitles.outputpath immediately after
init_project() so ALL intermediate saves (SaveProject after each
batch) go to the user's desired output location. Without this,
PySubtrans defaults to writing in the input directory — either as
*.translated.srt (new project) or *.{target_lang}.srt (resumed from
.subtrans via UpdateOutputPath) — which can overwrite existing
translations produced by a different provider. Resume capability
via .subtrans is fully preserved. (#259)
- test_semantic_diag.py: standalone diagnostic script for #267 semantic
scene detection hang. Reproduces the exact librosa feature extraction
steps outside WhisperJAV's subprocess architecture to isolate whether
a hang is system-level (librosa/numpy/BLAS) or subprocess-specific.
Accepts a media file as CLI argument, prints step-by-step progress
with timing at each stage.
- Issue tracker rev22: #267 diagnostic log analyzed (hang at
librosa.feature.mfcc, all versions compatible), standalone test sent.
#259 new bug (temp file overwrite), fix coded. #272 fix coded,
awaiting confirmation. #271 awaiting model info.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@@ -59,14 +58,14 @@ These fixes are committed to main after v1.8.10 was tagged. Colab/pip users get
59
58
60
59
| # | Title | Reporter | Status | Notes |
61
60
|---|-------|----------|--------|-------|
62
-
|**#272**| XXL ignores --pass2-model large-v2 flag | TinyRick1489 |`NEEDS RESPONSE`|**NEW.** Confirmed bug with debug log. WhisperJAV logs model=large-v2 but XXL subprocess loads large-v3. Flag not forwarded to XXL command line. |
63
-
|**#271**| Ollama "Could Not Create Optimized Variant" | justantopair-ai |`NEEDS RESPONSE`|**NEW.** qwen3:14b error. Also reports gist timeout fetching instructions_standard.txt, plot summary cutoff in GUI, translation errors. 3 screenshots. |
64
-
|**#267**| Stuck on streaming features (Qwen+Semantic) | OrangeFarmHorse |`NEEDS FOLLOW-UP`|**Debug logs arrived.** All 3 test runs (short clip, long file, balanced pipeline) hang at exact same point: "[1/5] Streaming features (60s chunks)..." inside SemanticAudioClustering v7.0.0. File length irrelevant. RTX 5060 Ti (Blackwell). |
65
-
|**#265**| Hallucination in Chinese translation (large-v2) | yangming2027 |`NEEDS FOLLOW-UP`|User provided 3 concrete post-processing suggestions: (1) trailing period inconsistency, (2) leading dash removal, (3) long subtitle line wrapping. Sample SRT files attached. Also requesting merge module (#230). |
61
+
|**#272**| XXL ignores --pass2-model large-v2 flag | TinyRick1489 |`FIX CODED`|Fix on main (`e50df15`). Responded with upgrade instructions. Awaiting user confirmation. |
62
+
|**#271**| Ollama "Could Not Create Optimized Variant" | justantopair-ai |`AWAITING REPLY`|Asked which model. Non-fatal variant warning + translation format mismatch — likely model not suited for translation. |
63
+
|**#267**| Stuck on streaming features (Qwen+Semantic) | OrangeFarmHorse |`AWAITING REPLY`|Hang pinpointed at `librosa.feature.mfcc()`. Standalone test script (`test_semantic_diag.py`) sent to user to isolate system vs subprocess issue. Waiting on results. |
64
+
|**#265**| Hallucination in Chinese translation (large-v2) | yangming2027 |`NEEDS FOLLOW-UP`|3 post-processing suggestions with sample SRTs: trailing period, leading dash, line wrapping. Also shared Subtitle Edit multi-replace XML. |
66
65
|**#263**| GPU not utilized / stuck at VAD | herlong6529424-dot |`AWAITING REPLY`| RTX 3070 8GB. Process hangs during Silero VAD. Asked for debug log + short file test. No response yet. |
67
66
|**#261**| Network check: unknown url type https | henry99a |`AWAITING REPLY`| NSIS SSL context issue. User found workaround (run from cmd). |
|**#259**| Local Translation Issues (hotfix2) | destinyawaits |`RESPONDED`|Responded with v1.8.10 Ollama upgrade instructions. |
68
+
|**#259**| Local Translation Issues (hotfix2) | destinyawaits |`NEEDS FOLLOW-UP`|**Ollama confirmed working.** New bug: temp file overwrites existing English subtitles in same folder — save path not respected until task complete. |
All fixes are on main and accessible via `pip install --no-deps --upgrade`. No blocking bugs requiring a versioned hotfix release. The #272 XXL fix and #267 diagnostics are already available to pip/Colab users.
164
159
165
-
| Bug | Issue | Severity | Hotfix? |
166
-
|---|---|---|---|
167
-
| XXL ignores --pass2-model flag |#272| MEDIUM |**Yes — if fix is small.** Confirmed bug, user on Kaggle CLI. |
168
-
| Semantic scene detect hangs on Blackwell GPUs |#267| HIGH |**No.** Needs investigation first. Workaround exists (auditok). |
**Recommendation**: Investigate #272 first. If the fix is a one-liner (model flag not forwarded to XXL subprocess args), it's a good hotfix candidate alongside the already-pushed fixes.
160
+
A hotfix would only be needed if:
161
+
- A blocking bug is found that affects .exe installer users specifically
162
+
- The #267 semantic hang turns out to be a code bug (currently investigating)
172
163
173
164
### v1.9.0 — Next Major Release
174
165
@@ -179,6 +170,7 @@ Colab/pip users get these automatically. .exe installer users get them via `whis
179
170
|**P0**| Ollama full migration (deprecate llama-cpp-python) |#132, #233, #255, #259| Large | Remove ~1500 LOC. llama-cpp already broken on Colab. |
180
171
|**P0**| Chinese UI (partial i18n) |#175, #180| Medium | 40%+ issues are Chinese. Biggest support burden reducer. |
181
172
|**P0**| Unified CLI override layer |#269| Small | Fix architectural divergence between standard/ensemble override paths. |
173
+
|**P1**| Translation temp file overwrite fix |#259| Small | Save path not respected until task complete — overwrites existing SRTs. |
182
174
|**P1**| Post-translation hallucination filter (Chinese) |#265| Medium | User contributed 130-rule XML. Automate as post-process option. |
183
175
|**P1**| Post-processing polish (trailing period, leading dash, line wrapping) |#265| Small | 3 concrete suggestions from yangming2027 with sample SRTs. |
184
176
|**P1**| Standalone merge CLI |#230| Medium |`whisperjav-merge`. Active demand (yangming2027, weifu8435). |
@@ -187,7 +179,7 @@ Colab/pip users get these automatically. .exe installer users get them via `whis
187
179
|**P1**| AMD ROCm support |#142, #114, #239| Medium | Document + test. |
188
180
|**P1**| Additional translation targets |#268| Small | Thai, Korean requested. May be PySubtrans config only. |
189
181
|**P1**| Full dual-track enhance-for-VAD | — | Medium | ASR module separate VAD/ASR audio paths. |
190
-
|**P2**| Semantic scene detection hanging investigation|#267| Medium |Hangs at "[1/5] Streaming features" on Blackwell GPU. Debug logs available. Needs code-level investigation. |
182
+
|**P2**| Semantic scene detection: librosa.feature.mfcc hang|#267| Medium |Diagnostic log pinpointed hang at first `librosa.feature.mfcc()` call. Versions all compatible. Suspected numba JIT issue — needs further investigation. |
191
183
|**P2**| GUI settings persistence |#96| Medium | Long-standing request. |
- Audio content (auditok processes same files fine)
240
+
- Resampling (native SR = target SR, no resample triggered)
236
241
237
-
**Hypothesis (unverified)**: The feature streaming step may use a GPU operation (torch audio feature extraction) that is incompatible with the Blackwell/RTX 50 series drivers or CUDA version. The hang occurs inside a library call, not in WhisperJAV code.
242
+
**What remains**: The hang is at the first `librosa.feature.mfcc()` call. This function uses numpy/scipy for FFT and DCT — no numba directly. However, librosa may trigger numba JIT compilation through internal utility imports. numba 0.64.0 is very recent and could have platform-specific issues.
238
243
239
-
**Next step**: Read the semantic scene detection source code to identify what "[1/5] Streaming features" does — what library calls, what GPU operations. Determine if there's a known Blackwell compatibility issue.
244
+
**Next diagnostic step**: Ask user to test with `NUMBA_DISABLE_JIT=1` environment variable to confirm/eliminate numba as the cause.
240
245
241
246
### Silero VAD Hanging (#263)
242
247
243
-
**Symptom**: Process hangs during Silero VAD initialization/inference. Log shows "Starting speech segmentation with: silero-v4.0" with no "complete" message. User has RTX 3070 8GB, processing a 2.4h file (314 scenes).
244
-
**Status**: Waiting on user debug log + short file test. No response since 2026-03-31 request.
248
+
**Symptom**: Process hangs during Silero VAD. Log shows "Starting speech segmentation with: silero-v4.0" with no "complete" message. RTX 3070 8GB.
249
+
**Status**: Waiting on user debug log + short file test. No response since 2026-03-31.
245
250
246
-
### XXL Model Flag Ignored (#272) — CONFIRMED BUG
251
+
### Translation Temp File Overwrite (#259) — NEW BUG
247
252
248
-
**Symptom**: `--pass2-model large-v2` is logged correctly by WhisperJAV but not forwarded to the Faster-Whisper-XXL subprocess. XXL loads large-v3 by default.
249
-
**Evidence**: Debug log shows `[Worker 3816] Pass 2: BYOP XXL (exe=..., model=large-v2)` but XXL subprocess output says `'large-v3' model may produce worse results`.
250
-
**Root cause**: The `--model` flag is likely missing from the XXL subprocess command construction in `pass_worker.py` or `xxl_runner.py`.
253
+
**Symptom**: During AI SRT Translate, the temp file is created in the same folder as the input SRT and overwrites existing English subtitles generated earlier (e.g., by DeepSeek). The user-specified save path is not used until the task completes.
254
+
**Reporter**: destinyawaits (also confirmed Ollama working after v1.8.10 upgrade).
255
+
**Status**: Needs investigation — is this PySubtrans behavior or WhisperJAV's translation wrapper?
251
256
252
-
### Hallucination in Chinese Translation (#265)
257
+
### Ollama Model Suitability (#271)
253
258
254
-
**Symptom**: After Japanese transcription + DeepSeek translation to Chinese, hallucination phrases appear in Chinese output.
255
-
**Assessment**: Two contributing factors: (1) hallucination filter was not bundled in wheel — fix committed; (2) no post-translation sanitization step exists. User provided 3 concrete post-processing suggestions with sample SRTs.
259
+
**Symptom**: Translation format errors with Ollama. Model output doesn't match PySubtrans expected format.
260
+
**Assessment**: Likely model not suited for structured translation tasks. Asked user which model they're using. Non-fatal "optimized variant" warning is separate (Ollama rejected Modelfile create with HTTP 400).
|**2026-03-31**|**rev21.** 54→55 open. New: #271 (Ollama variant error), #272 (XXL model flag bug). Closed: #269. #267 debug logs arrived — hangs at "[1/5] Streaming features" regardless of file length/pipeline, Blackwell GPU. #272 confirmed with debug log — model flag not forwarded to XXL subprocess. |
283
-
| 2026-03-31 |**rev20.** 55→54 open. New: #265, #267, #268, #269. Closed: #218, #221, #236, #244. All fixes pushed to main (4 commits). Colab tested. All NEEDS RESPONSE cleared. |
288
+
|**2026-04-01**|**rev22.** 55 open (unchanged). Committed `e50df15`: XXL model flag fix (#272) + semantic diagnostics v7.1.0 (#267). Responded to #271, #272, #267. #267 diagnostic log arrived — hang pinpointed at `librosa.feature.mfcc()` first call. All library versions compatible. Next step: test NUMBA_DISABLE_JIT=1. #259 user confirmed Ollama working + reported new temp file overwrite bug. All NEEDS RESPONSE cleared. |
0 commit comments