| ASR |
Automatic Speech Recognition |
| Parakeet |
NVIDIA NeMo ASR model family; Sbobinator uses TDT 0.6B v3 |
| NeMo |
NVIDIA NeMo Toolkit — framework for speech/LLM models |
| mT5 |
Multilingual T5 — Google seq2seq model for abstractive summarization |
| RTF |
Real-Time Factor — processing_time / audio_duration; <1 = faster than realtime |
| SRT |
SubRip — subtitle format with timestamps |
| FIFO |
First In, First Out — job queue order |
| WAL |
Write-Ahead Logging — SQLite mode for read/write concurrency |
| Worker |
Separate process that consumes the job queue |
| Pipeline |
Sequence extract → transcribe → summarize → export for one job |
| Chunking |
Splitting long audio into overlapping segments for ASR |
| Extractive |
Summary that selects sentences from the original text |
| Abstractive |
Summary that generates new text (mT5) |
| Stem |
Filename without extension |
| SBOBINATOR_DATA |
Environment variable for data root (data/ or /data in Docker) |
| NEMO_CACHE_DIR |
Environment variable for NeMo/mT5 model directory |
| Streamlit |
Python framework for interactive web UI |
| Typer |
CLI library based on Click |
| ffmpeg |
Tool for audio conversion and extraction from video |
| ffprobe |
Tool for media metadata (duration, codec) |
| Headless |
Execution without a graphical interface (CLI/worker only) |
| Orphan job |
Job left in running after worker crash — recovered on restart |