WhisperJAV

A subtitle generator for Japanese Adult Videos. Free, runs on your own machine, no cloud upload of your media.

Documentation: English | 简体中文

The idea

Speech recognition models like Whisper are trained on clean, curated speech. JAV audio is the opposite of that, and the mismatch breaks them in specific, well-understood ways:

1. The acoustic profile. JAV audio has a low signal-to-noise ratio and a high density of non-verbal vocalisations — breathing, gasps, moans — whose spectra often mimic real Japanese syllables (e.g. fu), tricking the model into hearing words where none exist. Add extreme volume swings (whispers to screams) and theatrical role language (yakuwarigo) absent from training corpora, and the model's assumptions stop holding.

2. Long-form drift and hallucination. These are feature-length recordings, not 30-second clips. Over long stretches of ambiguous audio — silence, rhythmic breathing — the model's attention collapses and it fills the void with repeated or invented text. This is a documented failure mode of Whisper-family models [3, 4, 5].

3. The pre-processing paradox. Intuition says "denoise first". In practice, blanket denoising and vocal isolation can strip the high-frequency detail the model needs to tell consonants apart, making things worse. Fine-tuning on JAV data has its own trap: good datasets are scarce, so fine-tuned models tend to overfit and become hit-or-miss.

WhisperJAV is built around these three failure points rather than around any single model:

• Scene-based segmentation — split the audio at natural acoustic boundaries so the model always processes a coherent environment, never a mixed stream [1, 2].
• VAD clamping — detect where speech actually is, and feed the model only that, with measured padding. This is the main defence against hallucination on non-speech.
• Defensive decoding and post-processing — tuned confidence thresholds discard low-quality output, and Japanese-aware filters clean what remains.

None of this is magic; it is careful plumbing around known model weaknesses, and the defaults are tuned against ground-truth benchmarks. Results still vary with source audio quality.

Quick start

GUI (recommended): launch WhisperJAV from the desktop shortcut (Windows installer) or run:

whisperjav-gui

Add files, pick a mode, click Start. Subtitles land next to your video as .srt.

Command line:

whisperjav video.mp4                                        # defaults
whisperjav video.mp4 --mode balanced --sensitivity aggressive
whisperjav /path/to/folder --output-dir ./subtitles         # whole folder

Any input FFmpeg can read works: MP4, MKV, AVI, WMV, MP3, WAV, FLAC, and so on. Output is SRT (default), WebVTT, or both (--output-format both).

New in v1.9.5

A small fix release on top of 1.9.4. On 1.9.3 or 1.9.4, upgrade with one command, whisperjav-upgrade --wheel-only, because 1.9.5 changes no library underneath WhisperJAV. Full detail: release notes.

• Saved output folder (GUI): with Remember settings on, an empty saved output folder sent subtitles into WhisperJAV's program folder. They now go next to the video (#454).
• Ollama translations: repeated <summary> and <scene> blocks in a local model's reply are now removed before the reply is read (#455, PR #443).
• DeepSeek: the default model is now deepseek-flash. DeepSeek still accepts the old name, deepseek-v4-flash (PR #457).
• Ensemble defaults: pass 1 is Qwen3-ASR + FireRedVAD and pass 2 is Anime-Whisper + TEN VAD, both Aggressive. Pass 2 stays off until you tick it. Saved settings are kept.
• Colab: the notebook no longer replaces Colab's PyTorch. It adds the CUDA 12 libraries that faster-whisper needs (#456, #447).

If you run WhisperJAV from scripts: in a two-pass run, a pass 1 on Qwen3-ASR or a pass 2 on anime-whisper now defaults to Aggressive with FireRedVAD (pass 1) or TEN (pass 2). In 1.9.4 it was Balanced with WhisperSeg. A pass 1 on Qwen3-ASR stops before transcribing if FireRedVAD is not installed. See Changes.

New in v1.9.4

Bug fixes, hardening, and better subtitle timing for ChronosJAV. On 1.9.3, upgrade with one command — whisperjav-upgrade --wheel-only — because 1.9.4 changes no library underneath WhisperJAV. Full detail, including every change and who reported it: release notes.

Subtitles stay in time in films with holes in their audio. Some downloaded films have hidden holes in their audio track. In those files every subtitle after the first hole came out early — by up to 10 seconds in the films we measured. WhisperJAV now fills the holes with silence while extracting the audio, so subtitle times stay in step with the video. Every file is checked as it is extracted. A damaged file is still transcribed, and the run summary lists it as suspect, with the reason. To stop such files instead, use --fail-on suspect, or tick Treat 'suspect' files as failures in the GUI.

An attempt to improve ChronosJAV timing (#433, #421, #417, #437, #427). For Qwen3-ASR and anime-whisper, the longest piece of speech sent to the model is now 4 seconds at every sensitivity, with the WhisperSeg, TEN and FireRedVAD speech segmenters, and anime-whisper has new settings for where long pieces are cut. TEN now cuts long speech at the quietest point and never past the limit. On seven drama scenes with reference subtitles, measured with WhisperSeg, the typical gap between a line's end and the reference end shrank by 10 to 24 % (Qwen3-ASR measured at balanced only). The text error changed by between 1.5 % better and 0.5 % worse. These are drama scenes, not JAV, and lines still end somewhat late — please tell us how it does on your films. Qwen3-ASR with WhisperSeg no longer adds 100 ms of padding after each piece of speech. anime-whisper now also hears 200 ms of silence before each piece of audio, except with the Silero speech segmenters, where it is off: Silence Before Each Window (ms), or --qwen-leading-silence (0 turns it off).

If you run WhisperJAV from scripts: --fail-on suspect now also stops files with damaged audio; a ForcedAligner that cannot be loaded now leaves the file suspect and exits 0 instead of failing the job; a custom translation server now keeps its own temperature unless you pass --temperature. See What will break.

Also here: Gemini's default model is now gemini-3.6-flash; translation choices are remembered between launches; the command line finds the bundled FFmpeg on Windows; on Balanced and Faster, a speech piece of a few milliseconds at the edge of a scene, which could crash the run, is now joined to its neighbour; and the Windows installer downloads PyTorch again if it does not load after install.

New in v1.9.3

More bug fixes and overall improvements. 1.9.3 builds on 1.9.2: where that release was about performance and the Balanced pipeline, this one is about fixing bugs and making small improvements across the features. Full detail, including every fix and who reported it: release notes.

The one to know about: a major bug in the speech enhancers. "Enhance for VAD only" did not do what its name says — on Fidelity it enhanced both tracks, so the recogniser heard the cleaned-up audio too, which is the very thing the option exists to avoid. It now genuinely splits them. The control was also invisible in the GUI unless you re-picked the enhancer by hand, and FFmpeg DSP could not be chosen from the GUI at all.

The recommended recipe — more accurate subtitles and more accurate timing, for a little more time. Choose a speech enhancer: htdemucs (new here; it comes with a fresh install — upgraders, see the release notes), or any of the stable enhancers that suits your machine. Then turn on Enhance for VAD only. The enhanced vocal stem drives the voice detection, so the speech is segmented on clean audio, while the recogniser transcribes the original, untouched audio. Preparing one track and recognising the other is what improves the accuracy. The maintainer's own favourite is **Qwen ASR + htdemucs + Enhance-for-VAD-only

• TEN VAD**:

whisperjav video.mkv --ensemble --pass1-pipeline qwen --pass1-model Qwen/Qwen3-ASR-1.7B \
  --pass1-speech-enhancer htdemucs --pass1-enhance-for-vad --pass1-speech-segmenter ten

If you run WhisperJAV from scripts: --mode balanced --enhance-for-vad now exits 2 with an explanation instead of being accepted and ignored, and so does a Balanced pass in a two-pass run. See What will break.

Also here: the installer shows what it is doing during the PyTorch download instead of looking frozen, and no longer writes settings into your global Git configuration; French translation works and Italian, Thai and Korean are added; a failed translation names the right server's log instead of telling everyone to check Ollama's.

New in v1.9.2

Full detail, including every bug fixed and who reported it: release notes.

Every run now tells you whether it worked. It ends with a table, one line per file:

STATE    MILEAGE        FILE
done     ok 100%        movie1.mp4 -> movie1.ja.srt (1961 cue(s))
suspect  low 6%         movie2.mp4 -> movie2.ja.srt (82 cue(s)); subtitles stop at 400s of 7000s (5.7% of the file); 82 cue(s) produced
empty    not assessed   movie3.mp4 -> movie3.ja.srt; movie3.ja.srt contains no subtitles

(The framing rules and the totals line are omitted here.)

done · empty · suspect · failed · skipped mean the same thing in every mode, and the exit status follows them: 1 if any file failed, or if the run did not finish (interrupted, or stopped by an error), 0 otherwise. Zero subtitles is an observation, not a failure — silence and music both end there — so empty and suspect exit 0 unless you ask otherwise with --fail-on empty / --fail-on suspect, or the two matching checkboxes on the GUI's Transcription Adv. Options tab. A whisperjav_run.json next to your subtitles carries the same result for scripts.

MILEAGE is how far into the file your subtitles reach — where the last one ends, as a share of the video's length. It is not how much of the video carries subtitles. --min-coverage sets how low that can go before a file is called suspect (a quarter, by default).

It warns you when your audio looks difficult. After working out the scenes and before the long part, a run on Balanced, Fidelity or --mode qwen says how much of the file is speech, and flags the file as unusually quiet, or names the scenes where speech sits close in level to everything around it. If your audio looks normal you get one line and nothing else.

Other things you can now do:

Changes you may notice: Aggressive is much faster on Balanced and no longer decodes a passage twice, so run times are predictable and the same file gives the same subtitles every time. Scene detection changed default (above), so your scenes — and your subtitles — will differ from 1.9.0/1.9.1 output. --no-vad is removed; --speech-segmenter none does the same thing on the modes that still take one. On a machine with no usable GPU a script now needs --accept-cpu-mode or --device cpu, where it used to continue by itself after a countdown.

How a video becomes subtitles

Every pipeline follows the same overall shape; modes differ in which components they use and how aggressively they are tuned.

flowchart LR
    A[Audio<br/>extraction] --> B[Scene<br/>detection]
    B --> C[Speech<br/>enhancement<br/><i>optional</i>]
    C --> D[Speech<br/>segmentation<br/>VAD]
    D --> E[ASR<br/>model]
    E --> F[Post-<br/>processing]
    F --> G[.srt]

• Scene detection uses the media characteristics to cut by predicted scenes. This allows the downstream VAD and ASR to receive chunks which have similar DNA.
• Speech enhancement (off by default) can clean audio per-scene — used surgically, per the pre-processing paradox above.
• Speech segmentation (VAD) finds where speech actually is inside each scene. This choice matters more than most settings: it decides what the model hears and, in the modern pipelines, where your subtitle timestamps come from.
• The ASR model turns speech into text.
• Post-processing is the Japanese-specific cleanup pass:
  • Sentence regrouping aware of ending particles (ね, よ, わ, の), aizuchi (うん, はい), and dialect patterns (Kansai-ben and others)
  • Hallucination and repetition removal
  • Sound-only line removal — subtitle lines that are purely moans/breathing kana are dropped (real dialogue is protected by an evidence check)
  • Timing repair — a subtitle whose duration is absurdly long for its text gets its start pulled in (the end stays put); the console reports how many lines were retimed
  • Scene-boundary overlap resolution

Each stage has several interchangeable providers — the full menu, with strengths and weaknesses, is in Mix-and-match strategies below.

Processing modes

Sensitivity applies to every mode: conservative (fewer false positives, good for noisy content) · balanced · aggressive (catches more quiet dialogue; good for whisper/ASMR content — and the tuning target of most of our benchmark work).

ChronosJAV

Some of the best recognizers for this domain (anime-whisper, Qwen3-ASR and its Japanese finetunes) don't produce reliable timestamps on their own. ChronosJAV runs text generation and timing as separate stages: the VAD provides the time skeleton, the model provides the words. Since v1.9, timestamps come from the VAD frames by default (no aligner model loaded, ~1 GB less VRAM); a Qwen forced-aligner mode remains available in the settings for word-level alignment.

The same decoupled design is why new models can be added without rebuilding the pipeline — anything that turns audio into text can be slotted in.

Two-pass ensemble

Different pipelines miss different lines. Ensemble mode runs your file through two pipelines and merges the results.

The v1.9.5 default pairing: pass 1 = Qwen3-ASR with FireRedVAD, pass 2 = anime-whisper with TEN VAD, both on Aggressive — two different recognizers and two different VADs, so their blind spots don't overlap. In the GUI, pass 2 runs once you tick it.

whisperjav video.mp4 --ensemble \
    --pass1-pipeline qwen --pass2-pipeline balanced \
    --merge-strategy pass1_primary

• Merge strategies: pass1_primary / pass2_primary (one pass leads, the other fills gaps), smart_merge, full_merge, pass1_overlap / pass2_overlap, longest
• Presets: save and reload named ensemble configurations from the GUI
• Serial mode (--ensemble-serial): finish each file completely before starting the next, so results appear as they're done
• Bring your own pass 1, pass 2 or both: PurfView's Faster-Whisper XXL (--pass1-pipeline xxl and/or --pass2-pipeline xxl, with --xxl-exe ...; the model with --pass1-model / --pass2-model, default large-v2)

Mix-and-match strategies

The Ensemble tab is a mixing desk. Each pass is a free combination of five choices — pipeline × scene detection × audio pre-processing × speech segmentation × ASR model — and the two-pass design is the sixth dimension. The defaults are benchmark-tuned, so you never have to touch any of this; but audio varies a lot, and one deliberate substitution is often worth the experiment. The golden rule: change one thing at a time, so you know what caused the difference.

Pipeline

The recipe that ties the other choices together.

Naming note (CLI users). ChronosJAV is one pipeline with interchangeable recognizers, so qwen and anime-whisper are two backends of it rather than two separate pipelines. The GUI lists them side by side in the pipeline dropdown; on the command line both live under qwen, and the backend is chosen inside the params:

# Qwen3-ASR (the default backend)
--pass1-pipeline qwen

# anime-whisper, same pipeline, different recognizer
--pass1-pipeline qwen --pass1-qwen-params '{"generator_backend": "anime-whisper"}'

The usable values for generator_backend are qwen3 and anime-whisper (a third, cohere, exists in the code but is not currently enabled). Note also that litagin/anime-whisper can be loaded through the transformers pipeline via --hf-model-id — but that route uses HF's own chunking and does not produce usable per-line timing, so prefer the ChronosJAV route above.

Scene detection

Where the long file gets cut into workable pieces.

Audio pre-processing (speech enhancement)

Off by default — remember the pre-processing paradox. The "Enhance for VAD only" checkbox is the safest way to use these: the cleaned audio guides speech detection while the model still hears the original.

Speech segmentation (VAD)

Decides what the model hears — and in ChronosJAV pipelines, where your timestamps come from. Probably the highest-leverage swap on this list.

Not on Balanced, since v1.9.2. The Balanced pipeline uses Faster-Whisper's own voice detection and accepts none of the backends below — --speech-segmenter with --mode balanced stops the run with an error, as do --max-group-duration and --chunk-threshold. What you choose there instead is which Silero build that detection runs: --vad-version 3.1 | 4.0 | 6.2 (default 4.0), or the Speech Segmenter column of a Balanced pass in the Ensemble Mode tab. Every other pipeline takes the table below unchanged.

ASR engine and model

The two-pass dimension

Everything above multiplies: two passes means two full recipes, then a merge. What makes a good pair is diversity — different recognizers and different VADs, so the passes fail in different places and the merge covers both.

A few known-good recipes:

Merge strategies. Seven are available (--merge-strategy):

Save anything that works as a preset so it's one click next time.

AI translation

Generate and translate in one go, or translate subtitles you already have:

whisperjav video.mp4 --translate                      # transcribe + translate
whisperjav-translate -i subtitles.srt --provider ollama

Interrupted translations resume where they left off — just run the same command again.

The GUI

Four tabs:

1. Transcribe — files, mode, sensitivity, language
2. Advanced options — output format, scene detection, model override, debug
3. Ensemble — the two-pass grid: per-pass pipeline, sensitivity, scene detector, enhancer, VAD, and model, plus a Customize dialog exposing each backend's tunable parameters, and preset save/load
4. AI SRT Translate — translate existing subtitle files

Sensible defaults everywhere: if you never open a Customize dialog, you get the benchmark-tuned configuration.

Which mode for which content

These are starting points, not rules — see Mix-and-match strategies for how to adapt them.

Installation

Already on 1.10? Windows installer: upgrade with whisperjav-upgrade; it keeps the accelerator of your install. Source install: git fetch --tags, git checkout the new release tag, and run the installer again with the same Python environment active. Rollback is available: whisperjav-upgrade --rollback. On 1.9.x? Install 1.10.0 fresh: it needs Python 3.13, so whisperjav-upgrade cannot take 1.9.x to it.

Each install is made for one accelerator, chosen when you install: NVIDIA GPU (CUDA), GPU through Vulkan (AMD, Intel or older NVIDIA card, Windows), Apple GPU (Metal), your own PyTorch, or CPU only. See Accelerated or not.

Windows — standalone installer (recommended)

No Python knowledge needed.

1. Download the .exe from Releases
2. Run it — no admin rights required (installs to %LOCALAPPDATA%\WhisperJAV)
3. Wait 10–30 minutes while it sets up Python, PyTorch, FFmpeg and dependencies. With a supported NVIDIA GPU (GTX 16 / RTX 20 or newer, driver 580+) it installs for the GPU by itself; otherwise a window asks what the install should be made for
4. Launch from the desktop shortcut. First transcription downloads models (~3 GB)

Google Colab / Kaggle

Not supported by 1.10.0, which is made and tested for installs on your own computer only. The Colab expert notebook (badge at the top) keeps installing 1.9.5.

git clone --branch v1.10.0 https://github.com/meizhong986/whisperjav.git
cd whisperjav
installer\install_windows.bat                    :: a supported NVIDIA GPU is used; otherwise it asks
installer\install_windows.bat --accelerator cpu  :: or choose: nvidia, vulkan, own-pytorch, cpu

xcode-select --install
brew install python@3.13 ffmpeg portaudio git

git clone --branch v1.10.0 https://github.com/meizhong986/whisperjav.git
cd whisperjav
python3.13 -m venv ~/venvs/whisperjav && source ~/venvs/whisperjav/bin/activate
chmod +x installer/install_mac.sh && ./installer/install_mac.sh

sudo apt-get install -y python3 python3-pip python3-venv python3-dev \
    build-essential ffmpeg git libsndfile1 libsndfile1-dev

git clone --branch v1.10.0 https://github.com/meizhong986/whisperjav.git
cd whisperjav
chmod +x installer/install_linux.sh && ./installer/install_linux.sh

python install.py --accelerator own-pytorch

System requirements

Rough speed per hour of video: RTX GPU 5–10 min · Apple Silicon 8–15 min · CPU 30–60 min.

Without a supported NVIDIA GPU

Choose what the install is made for when you install — GPU through Vulkan (Windows: AMD, Intel or older NVIDIA card; the GGUF pipelines use it, experimental), your own PyTorch, or CPU only. Every run then follows that choice; nothing needs ticking. At the start of each run WhisperJAV says where each part runs. --device cpu runs one job on the CPU. Details: Accelerated or not.

Troubleshooting

• The Windows installer stopped and said the installation failed. Since v1.9.2 it checks that what WhisperJAV needs to run actually imports, and stops rather than leaving you a shortcut to a broken install. It names what failed and writes INSTALLATION_FAILED_v<version>.txt next to install_log_v<version>.txt in the installation folder. Run the installer again first — the usual cause is a download that did not finish. If it fails twice, attach both files to an issue.
• "FFmpeg not found" — install FFmpeg and add it to PATH.
• Very slow — run whisperjav --check, or read the box at the start of a run: they say where each part runs. On a CPU-only install everything runs on the CPU, several times slower than on a GPU.
• model.bin error in faster mode — enable Windows Developer Mode (or run once as admin), then delete the cached model folder under %USERPROFILE%\.cache\huggingface\hub.
• An empty or very short subtitle file, with the run reporting success. v1.9.2 fixes one cause of this on Balanced and now reports the outcome per file rather than always claiming success — see New in v1.9.2. If it still happens, that is worth a report.
• Anything else: open a GitHub issue with your system info and the console log. tools/whisperjav_env_report.py in this repository prints your package versions, GPU, driver and CUDA in one go — run it with the same Python that runs WhisperJAV (Windows users can right-click whisperjav_env_report.ps1). Logs and reproduction details make fixes much faster.

Contributing

Contributions are welcome — see CONTRIBUTING.md. Development setup:

git clone https://github.com/meizhong986/whisperjav.git
cd whisperjav
pip install -e ".[dev]"
python -m pytest tests/

License

MIT. See LICENSE.

References

1. Chen, Y., et al. (2025). "ChronusOmni: Improving Time Awareness of Omni Large Language Models." arXiv:2512.09841. (Inspiration for the ChronosJAV pipeline)
2. Bain, M., et al. (2023). "WhisperX: Time-Accurate Speech Transcription of Long-Form Audio." arXiv:2303.00747.
3. Wang, Y., et al. (2025). "Calm-Whisper: Reduce Whisper Hallucination On Non-Speech By Calming Crazy Heads Down." Interspeech 2025. arXiv:2505.12969.
4. Barański, M., et al. (2025). "Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio." arXiv:2501.11378.
5. Koenecke, A., et al. (2024). "Careless Whisper: Speech-to-Text Hallucination Harms." ACM FAccT 2024.
6. Roll, N., et al. (2025). "In-Context Learning Boosts Speech Recognition via Human-like Adaptation to Speakers and Language Varieties." arXiv:2505.14887.
7. Yang, X., et al. (2024). "PromptASR for Contextualized ASR with Controllable Style." ICASSP 2024. arXiv:2309.07414.

Acknowledgments

Built on the shoulders of: OpenAI Whisper · faster-whisper · stable-ts · Qwen3-ASR · anime-whisper · Kotoba-Whisper · HuggingFace Transformers · PySubtrans — and the testing community, whose feedback and bug reports shape every release.

Disclaimer

This tool generates accessibility subtitles. Users are responsible for compliance with applicable laws regarding the content they process.