Skip to content

Commit ce62256

Browse files
meizhong986meizhong986
committed
Merge v1.8.6 release
2 parents 552231c + 35dc1eb commit ce62256

37 files changed

Lines changed: 5002 additions & 334 deletions

‎README.md‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -8,7 +8,7 @@
88
<img src="https://kaggle.com/static/images/open-in-kaggle.svg" alt="Open In Kaggle"/>
99
</a>
1010
<br>
11-
<img src="https://img.shields.io/badge/version-1.8.5-blue.svg" alt="Version">
11+
<img src="https://img.shields.io/badge/version-1.8.6-blue.svg" alt="Version">
1212
<img src="https://img.shields.io/badge/python-3.10--3.12-green.svg" alt="Python">
1313
<img src="https://img.shields.io/badge/license-MIT-orange.svg" alt="License">
1414
</p>
@@ -637,6 +637,7 @@ MIT License. See [LICENSE](LICENSE) file.
637637
- In-Context Learning Boosts Speech Recognition." (2025). arXiv:2505.1
638638
- Koenecke, A., et al. (2024). "Careless Whisper: Speech-to-Text Hallucination Harms." ACM FAccT 2024.
639639
- Bain, M., et al. (2023). "WhisperX: Time-Accurate Speech Transcription of Long-Form Audio." arXiv:2303.00747.
640+
- Chen, Y., et al. (2025). "ChronusOmni: Improving Time Awareness of Omni Large Language Models." arXiv:2512.09841. *(Inspiration for the ChronosJAV pipeline)*
640641

641642

642643
## Acknowledgments
@@ -646,6 +647,7 @@ MIT License. See [LICENSE](LICENSE) file.
646647
- [stable-ts](https://github.com/jianfch/stable-ts) - Timestamp refinement
647648
- [faster-whisper](https://github.com/guillaumekln/faster-whisper) - Optimized CTranslate2 inference
648649
- [HuggingFace Transformers](https://github.com/huggingface/transformers) - Transformers pipeline backend
650+
- [Anime-Whisper](https://huggingface.co/litagin/anime-whisper) - Anime/JAV-tuned speech model (ChronosJAV pipeline)
649651
- [Kotoba-Whisper](https://huggingface.co/kotoba-tech/kotoba-whisper-v2.2) - Japanese-optimized Whisper model
650652
- [PySubtrans](https://github.com/machinewrapped/llm-subtrans) - AI-powered subtitle translation engine
651653
- The testing community for feedback and bug reports

‎RELEASE_NOTES_v1.8.6.md‎

Lines changed: 203 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,203 @@
1+
v1.8.6 introduces the **ChronosJAV pipeline** — a new dedicated pipeline for anime and JAV content with specialized speech models. This release also brings ensemble workflow improvements (presets, serial mode, settings persistence), new output formats, and fixes for translation on non-English systems.
2+
3+
---
4+
5+
### New: ChronosJAV Pipeline
6+
7+
The headline feature of v1.8.6. A new pipeline in the Ensemble tab built around models trained specifically for Japanese anime and adult content. Inspired by the temporal-awareness approach in [ChronusOmni](https://arxiv.org/abs/2512.09841) (Chen et al., 2025).
8+
9+
| Feature | Description |
10+
|---------|-------------|
11+
| **Anime-Whisper model** | New speech model fine-tuned for anime and JAV dialogue. Available in the GUI Ensemble tab under the ChronosJAV pipeline. Greedy decoding with TEN VAD segmentation for accurate timing. |
12+
| **Kotoba v2.0 and v2.1 models** | Two additional Japanese speech models available in the ChronosJAV model dropdown. Lighter weight (~2GB vs ~4GB) alternative to anime-whisper, based on the same Whisper large-v3 architecture. Kotoba v2.1 adds punctuation support. |
13+
| **Improved subtitle timing** | Retuned TEN VAD defaults for better segment boundaries. Eliminates oversized 15-36 second subtitle blocks and produces tighter, more natural subtitle timing while preserving sensitivity for short utterances and soft speech. |
14+
15+
---
16+
17+
### Key Features
18+
19+
| Feature | Description |
20+
|---------|-------------|
21+
| **Ensemble presets** | Save, load, and delete named ensemble configurations from the GUI. Reuse your tuned settings across sessions and across different pipeline combinations. Badge names for quick identification. |
22+
| **Settings persistence** | GUI pipeline and ensemble settings now survive application restarts. No more reconfiguring every session. Includes backup rotation and safe atomic writes. |
23+
| **VTT output format (#143)** | New output option: `--output-format srt/vtt/both` in CLI, or select from the dropdown in Advanced Options. Generates WebVTT subtitles for HTML5 video players. Works with all pipelines including ensemble. |
24+
| **Serial ensemble mode (#179)** | New `--ensemble-serial` option completes each file fully (Pass 1 → Pass 2 → Merge) before starting the next. See results as they finish instead of waiting for the entire batch. GUI checkbox in the merge strategy row. |
25+
26+
---
27+
28+
### Enhancements
29+
30+
| Enhancement | Description |
31+
|-------------|-------------|
32+
| **M4B audiobook support (#194)** | `.m4b` audiobook files can now be processed directly. Just drag and drop — FFmpeg handles the format natively. |
33+
| **Linux/macOS upgrade support** | `whisperjav-upgrade` now works on pip-based installations on Linux and macOS. Previously Windows-only. |
34+
| **Longest merge strategy** | New merge strategy option for ensemble mode, plus additional CLI quality knobs for advanced users. |
35+
36+
---
37+
38+
### Bug Fixes
39+
40+
| Issue | What was broken | What we fixed |
41+
|-------|-----------------|---------------|
42+
| #190 | Translation crashed on Chinese/Japanese Windows with a GBK codec error | Process-wide UTF-8 mode now covers all translation internals |
43+
| #188 | "Unknown translation provider: Gemini" on Linux pip installs — even with google-genai installed | Added missing `google-api-core` dependency (undeclared by upstream PySubtrans) |
44+
| #183 | "API token limit" and "No matches" errors during local LLM translation | Auto-detect model context window size and cap batch size accordingly |
45+
| #183 | `whisperjav-upgrade --wheel-only` showed "installation not found" on Linux | Cross-platform Python/pip detection for the upgrade tool |
46+
47+
### Breaking Changes
48+
49+
None.
50+
51+
---
52+
53+
## Installation Guide
54+
55+
### Upgrading from v1.8.5
56+
57+
Same dependency set — safe upgrade:
58+
59+
```bash
60+
whisperjav-upgrade
61+
```
62+
63+
Or wheel-only (code changes only, no dependency reinstall):
64+
65+
```bash
66+
whisperjav-upgrade --wheel-only
67+
```
68+
69+
**New:** `whisperjav-upgrade` now works on Linux and macOS pip-based installations.
70+
71+
---
72+
73+
### Windows -- Standalone Installer (Most Users)
74+
75+
The easiest way. No Python knowledge needed.
76+
77+
Recommended: Uninstall the old version first (Settings > Apps > WhisperJAV), then install fresh. Your models and output files are stored separately and won't be lost.
78+
79+
1. **Download:** WhisperJAV-1.8.6-Windows-x86_64.exe from below
80+
2. **Run the installer.** No admin rights required. Installs to `%LOCALAPPDATA%\WhisperJAV`.
81+
3. **Wait 10-20 minutes.** It downloads and configures Python, PyTorch, FFmpeg, and all dependencies.
82+
4. **Launch** from the Desktop shortcut.
83+
5. **First run** downloads models (~3 GB, another several minutes).
84+
85+
**GPU auto-detection:** The installer checks your NVIDIA driver version and picks the right PyTorch:
86+
- Driver 570+ gets CUDA 12.8 (optimal for RTX 20/30/40/50-series)
87+
- Driver 450-569 gets CUDA 11.8 (broad compatibility)
88+
- No NVIDIA GPU gets CPU-only mode
89+
90+
---
91+
92+
### Windows -- Source Install (Developers)
93+
94+
For people who manage their own Python environments.
95+
96+
**Prerequisites:** Python 3.10-3.12, Git, FFmpeg in PATH.
97+
98+
```batch
99+
git clone https://github.com/meizhong986/whisperjav.git
100+
cd whisperjav
101+
102+
:: Full automated install (auto-detects GPU)
103+
installer\install_windows.bat
104+
105+
:: Or with options:
106+
installer\install_windows.bat --cpu-only :: Force CPU
107+
installer\install_windows.bat --cuda118 :: Force CUDA 11.8
108+
installer\install_windows.bat --cuda128 :: Force CUDA 12.8
109+
installer\install_windows.bat --local-llm :: Include local LLM translation
110+
```
111+
112+
The installer runs in 5 phases: PyTorch first (with GPU detection), then scientific stack, Whisper packages, audio/CLI tools, and optional extras. This order matters -- PyTorch must be installed before anything that depends on it, or you end up with CPU-only wheels.
113+
114+
For the full walkthrough, see [docs/guides/installation_windows_python.md](docs/guides/installation_windows_python.md).
115+
116+
---
117+
118+
### macOS (Apple Silicon)
119+
120+
**Prerequisites:**
121+
```bash
122+
xcode-select --install # Xcode Command Line Tools
123+
brew install python@3.12 ffmpeg git # Or python@3.11
124+
```
125+
126+
**Install:**
127+
```bash
128+
git clone https://github.com/meizhong986/whisperjav.git
129+
cd whisperjav
130+
131+
# Create a virtual environment (required for Homebrew Python)
132+
python3 -m venv ~/venvs/whisperjav
133+
source ~/venvs/whisperjav/bin/activate
134+
135+
# Run the macOS installer
136+
chmod +x installer/install_mac.sh
137+
./installer/install_mac.sh
138+
```
139+
140+
**GPU acceleration:** Apple Silicon (M1/M2/M3/M4/M5) gets MPS acceleration automatically for Whisper pipelines. Use `--mode transformers` for best performance. The `balanced`, `fast`, and `faster` modes use CTranslate2 which doesn't support MPS, so those fall back to CPU.
141+
142+
**Qwen pipeline on Mac:** Currently runs on CPU only. The forced aligner doesn't detect MPS yet. This is a known limitation we plan to fix.
143+
144+
**Intel Macs:** CPU-only. No GPU acceleration available.
145+
146+
For the full walkthrough, see [docs/guides/installation_mac_apple_silicon.md](docs/guides/installation_mac_apple_silicon.md).
147+
148+
---
149+
150+
### Linux (Ubuntu, Debian, Fedora, Arch)
151+
152+
**1. Install system packages first** -- these can't come from pip:
153+
154+
```bash
155+
# Ubuntu / Debian
156+
sudo apt-get update
157+
sudo apt-get install -y python3 python3-pip python3-venv python3-dev \
158+
build-essential ffmpeg git libsndfile1 libsndfile1-dev
159+
160+
# Fedora / RHEL
161+
sudo dnf install -y python3 python3-pip python3-devel gcc gcc-c++ \
162+
ffmpeg git libsndfile libsndfile-devel
163+
164+
# Arch
165+
sudo pacman -S --noconfirm python python-pip base-devel ffmpeg git libsndfile
166+
```
167+
168+
For the GUI, you'll also need WebKit2GTK (`libwebkit2gtk-4.0-dev` on Ubuntu, `webkit2gtk4.0-devel` on Fedora).
169+
170+
**2. Install WhisperJAV:**
171+
172+
```bash
173+
git clone https://github.com/meizhong986/whisperjav.git
174+
cd whisperjav
175+
176+
# Recommended: use the install script
177+
chmod +x installer/install_linux.sh
178+
./installer/install_linux.sh
179+
180+
# With options:
181+
./installer/install_linux.sh --cpu-only
182+
./installer/install_linux.sh --local-llm
183+
```
184+
185+
**NVIDIA GPU:** You need the NVIDIA driver (450+ or 570+) but NOT the CUDA Toolkit -- PyTorch bundles its own CUDA runtime.
186+
187+
**PEP 668 note:** If your distro's Python is "externally managed" (Ubuntu 24.04+, Fedora 38+), you'll need a virtual environment. The install script detects this and tells you what to do.
188+
189+
For the full walkthrough including Colab/Kaggle setup, headless servers, and systemd services, see [docs/guides/installation_linux.md](docs/guides/installation_linux.md).
190+
191+
---
192+
193+
### Google Colab / Kaggle
194+
195+
The notebooks are not updated yet for this release.
196+
197+
---
198+
199+
## Full Changelog
200+
201+
**20 commits since v1.8.5**
202+
203+
[View full comparison](https://github.com/meizhong986/WhisperJAV/compare/v1.8.5...v1.8.6)

0 commit comments

Comments
 (0)