VinoForge — OpenVINO App Series

English | 简体中文

A monorepo for a series of OpenVINO-accelerated applications on Intel hardware.

Current apps:

App	Description
vino-sub-gen	Video/audio subtitle generation & translation (Whisper + Hy-MT2)

Each app lives under apps/ and shares common infrastructure from libs/ov_kit/.

Developed and measured on an Intel Core Ultra 7 265K + Arc B580.

Pipeline

video.mkv
    |  ffmpeg  ->  16 kHz mono wav
    v
Silero VAD  ->  speech segments
    |
    |  Whisper Large-v3 FP16 (OpenVINO, Arc GPU)
    v
source.srt
    |
    |  Hy-MT2 / Hunyuan Dense V1 (OpenVINO)
    v
translated.srt

Whisper's own translate task only ever emits English. For any other target language you must transcribe first and then translate — that is the reason for the two stages.

Repository layout

vino-forge/                     ← monorepo root
  libs/ov_kit/                  ← shared: paths, device enumeration
  apps/vino-sub-gen/            ← video subtitle generation & translation
    lib/                        ← ASR, translation, timestamp modules
    script/                     ← CLI entry points, batch driver
    gui/                        ← PySide6 GUI
    tool/model_conversion/      ← Hunyuan → OpenVINO export tools
    models/                     ← model cards (weights not committed)
  models/                       ← shared model weights (not committed)

Components (vino-sub-gen)

Path	Role
apps/vino-sub-gen/script/generate_and_translate.py	One-shot driver: video/audio → source SRT → translated SRT
apps/vino-sub-gen/script/batch_generate.py	Batch CLI: scan a folder and process each video
apps/vino-sub-gen/script/video_to_subtitle.py, apps/vino-sub-gen/script/translate_srt.py	CLI entry points for transcription and translation
apps/vino-sub-gen/script/whisper_timestamp_srt.py	Optional standalone CLI for Whisper internal timestamps
apps/vino-sub-gen/gui/gui_app.py	Optional PySide6 GUI for single-file and in-process batch processing
apps/vino-sub-gen/lib/	Reusable ASR, translation, and timestamp modules
apps/vino-sub-gen/tool/model_conversion/	Hunyuan → OpenVINO export (FP16 / INT4) + inference checks
models/	Model cards; download the weights into these folders

Requirements

Python 3.10, plus ffmpeg available on PATH (or point the FFMPEG_PATH environment variable at the binary).

pip install -r requirements.txt

Or assemble the environment manually:

conda create -n py310_openvino python=3.10
conda activate py310_openvino

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
pip install openvino nncf optimum[openvino] transformers soundfile silero-vad

nncf is only needed for INT4 quantization — skip it if you only download pre-converted models.

Getting started

1. Get the models

Both models are needed; the weights are not committed to this repository — download or convert them into the models/ folders so the script defaults find them:

Whisper Large-v3 FP16 (OpenVINO) — from OpenVINO/whisper-large-v3-fp16-ov (Apache-2.0):

hf download OpenVINO/whisper-large-v3-fp16-ov --local-dir models/whisper-large-v3-fp16-ov

Hy-MT2 — download the pre-converted INT4 7B from ModelScope: ahbencat/Hy-MT2-7B-ov-int4 (~4.2 GB):

modelscope download --model ahbencat/Hy-MT2-7B-ov-int4 --local_dir models/Hy-MT2-7B-ov-int4

Or convert it yourself (1.8B FP16, or tweaked INT4 settings) — see tool/model_conversion/README.md.

Any other location works too; point the scripts at it with --model-dir.

2. Generate subtitles

# One file: transcribe + translate (defaults: target language 中文)
python apps/vino-sub-gen/script/generate_and_translate.py "input.mkv"

# Known source language, English output
python apps/vino-sub-gen/script/generate_and_translate.py "input.mkv" --src-lang ja --tgt-lang English

# Transcription only, no translation
python apps/vino-sub-gen/script/generate_and_translate.py "input.mkv" --tgt-lang None

3. Or batch-process a folder

# Every video under ./videos (recursive, skips ones with an existing .srt)
python apps/vino-sub-gen/script/batch_generate.py "./videos"

# Preview what would run
python apps/vino-sub-gen/script/batch_generate.py "./videos" --dry-run

# Custom languages / devices
python apps/vino-sub-gen/script/batch_generate.py "./videos" --src-lang ja --tgt-lang English --asr-device GPU.1 --trans-device GPU

4. GUI

Install the optional GUI dependency and launch:

pip install PySide6
python apps/vino-sub-gen/gui/gui_app.py

Choose a media file or folder in the top row. Folder input runs the in-GUI batch pipeline, reusing the Whisper and Hy-MT2 models across files. Batch options include recursive scanning, skipping existing SRTs, and skipping folders marked -C / -CH. Stop cancels after the current model segment/translation block completes.

Individual stages

python apps/vino-sub-gen/script/video_to_subtitle.py "input.mkv" --task transcribe --language zh
python apps/vino-sub-gen/script/translate_srt.py input.srt --tgt-lang 中文

Reusable implementations live in apps/vino-sub-gen/lib/; the corresponding CLI entry points are in apps/vino-sub-gen/script/.

Performance

Measured on Intel Core Ultra 7 265K + Arc B580.

Transcription — Whisper Large-v3 FP16, 5 minutes of audio:

Mode	Device	batch_size	Time	Segments
VAD (default)	Arc B580 (GPU.1)	1	30.0 s	89
VAD	Arc B580	2	29.5 s	89
No VAD	Arc B580	2	27.7 s	22

A 2.5-hour movie completes in roughly 4.4 minutes and yields 333 subtitle entries.

Translation — Hy-MT2:

Model	Device	Speed	Footprint
1.8B FP16	Arc B580	~30+ tok/s	~3.5 GB
1.8B FP16	CPU	~10 tok/s	~3.5 GB
7B INT4	Arc B580	7.85 tok/s	~4.2 GB
7B INT4	CPU	2.45 tok/s	~4.2 GB

Devices

GPU.1 is the discrete Arc GPU and the default for ASR, GPU.0 is the integrated GPU, and CPU / AUTO work everywhere.

NPU does not work for either model. Intel's NPU requires fully static shapes, while both Whisper and the stateful-KV-cache LLM carry a dynamic sequence-length dimension. Use GPU or CPU.

Documentation

Document	Contents
apps/vino-sub-gen/tool/model_conversion/README.md	Hunyuan → OpenVINO conversion and inference guide
models/SILERO_VAD/README.md	Silero VAD model card and recovery steps
models/whisper-large-v3-fp16-ov/README.md	Whisper model card
models/Hy-MT2-7B-ov-int4/README.md	Hy-MT2 INT4 model card