pywhispercpp -> faster-whisper (VAD), subtitle length, language detection

[michal-lozowski]

SUMMARY:

Moved from pywhispercpp to faster-whisper: has VAD, has no built in output formatters txt/srt/vtt
Implemented: subtitle line length capping, multipoint language detection, CPU/GPU toggle via config, VAD visualization, output formatters

MODIFIED FILES:

Dockerfile

• Added CPU/GPU toggle blocks (comment/uncomment)

docker-compose.yaml

• Added CPU/GPU toggle blocks (comment/uncomment)

src/app.py

• New /vad endpoint for VAD segment data
• Formatters

src/core/Transcriber.py

• Rewritten for faster-whisper API
• Multi-point language detection (samples at 20%, 40%, 60%, 80%)
• VAD log capture from faster-whisper
• Word-level timestamps for subtitle splitting
• Configurable device/compute_type from env vars

src/core/TsApi.py

• WhisperModel init for faster-whisper
• Reads device/compute_type from env vars

NEW FILES:

src/utils/formatters.py

• Custom SRT/VTT/TXT/CSV/JSON formatters
• Replaces pywhispercpp's built-in formatters

src/utils/subtitle_splitter.py

• Splits long subtitles using word-level timestamps
• Splits at punctuation/spaces, max 80 chars per line
• Trims subtitles longer than 30s (whisper hallucination countermeasure crutch :)

src/utils/vad_capture.py

• Captures VAD segments from faster-whisper logs
• Generates VAD timeline CSVs

visualize_vad.py

• Generates VAD timeline as a graphic bar chart

QUICKSTART.md

• Curl commands
• CPU/GPU switching instructions