video-vec2wav2-tokenizer-2

video-vec2wav2-tokenizer-2 **Version 2** - continuation shard of the video-to-AI-dataset tokenizer project. Production-ready pipeline (Python package `video_vec2wav2_tokenizer`, CLI command `video2dataset`) that turns a folder of videos into clean **AI training datasets for speech recognition (ASR) and text-to-speech (TTS)**. **Video processing** — recursive scan of `mp4 / mkv / avi / mov / webm`, FFmpeg audio extraction to mono · 16 kHz · 16-bit PCM WAV. **Speech recognition** — faster-whisper, CPU & CUDA, automatic language detection, word-level timestamps. **Segmentation** — cut audio by transcript timestamps into `dataset/audio/000001.wav …`. **Dataset generation** — `metadata.csv`, `dataset.jsonl`, `tts_metadata.csv`. **Feature extraction** (optional) — streaming `features/train.bin` + `train.dat` with float32 samples, mel spectrograms, duration and sample rate. **Statistics** — `report.json` with totals, durations and language distribution. **Training** — `train_wav2vec2.py`: HuggingFace Wav2Vec2 CTC with resume, multi-GPU, mixed precision and checkpointing. **Performance** — multiprocessing, batch processing, tqdm progress bars and memory-efficient streaming that scales past 1 TB of source media. Installation Requires **Python 3.11+** and the **FFmpeg** binary on your `PATH`. Verify: Quick start Outputs land under `dataset/`: CLI Global flags (override `config.yaml`): `--config`, `--output-dir`, `--device {auto,cpu,cuda}`, `--whisper-model`, `--language`, `-v/--verbose`. Multi-GPU training Configuration (`config.yaml`) `speaker_mode` controls TTS speaker identification: a single fixed speaker, one per source video, or one per clip. Architecture Design notes: **Lazy heavy imports.** `faster-whisper`, `torch` and `transformers` are imported only when their command runs, so the core library and tests stay light and fast. **Streaming everywhere.** Transcription consumes whisper's generator; clips and manifest rows are produced per source video; the feature store