Skip to content

TODO.md

Latest commit

 

History

History
63 lines (53 loc) · 3.25 KB

TODO.md

File metadata and controls

63 lines (53 loc) · 3.25 KB
Raw

TODO & Roadmap

Current Status (2026-02-24)

Phase 1 (timestamps) and Phase 2 (diarization) are fully implemented and tested. Ready for production runs.

Completed

  • Plain text transcription of all 1,013 files (output/)
  • Phase 1: CTC timestamp extraction
  • Phase 1: 5-file and 20-file timestamp tests (20/20 passed)
  • Phase 2: Speaker diarization via pyannote.audio 3.1.1
  • Phase 2: 5-file and 20-file diarized tests (20/20 passed)
  • Patch scripts for pyannote + torchaudio nightly compatibility
  • HuggingFace gated model access (segmentation-3.0 license accepted)
  • Documentation (README, CHANGELOG, SETUP_GUIDE, DEVELOPMENT_NOTES)
  • Project code in local-copy-github for version control
  • Model backups to NAS (Parakeet 8GB + segmentation 5.7MB + wespeaker 26MB)
  • NFS migration — Spark switched from CIFS to NFS, fstab updated
  • All project files synced to NAS via NFS

Next Steps

Immediate — Full Production Run

  • Decision: Run timestamps-only (~1.5 hrs) or diarized (~53 hrs)?
    • Most videos in the corpus are solo — diarization adds minimal value for those
    • Consider: timestamps-only for all, then diarized for known multi-speaker videos
  • Run full production batch in tmux on Spark
  • Verify output quality on a sample of completed transcripts
  • Archive plain text output/ (v1) before overwriting

Short Term

  • GPU diarization: Monitor PyTorch nightly for sm_121 FFT kernel support
    • Check: python -c "import torch; t = torch.randn(100, device='cuda'); torch.fft.rfft(t)"
    • When this works, remove CPU fallback in DiarizationEngine for ~10x speedup
  • Selective diarization: Flag multi-speaker files and only diarize those
    • Could use a quick VAD pass to estimate speaker count before full diarization
  • Output post-processing: Consider sentence casing, punctuation restoration
    • Parakeet outputs lowercase only — a small model could add punctuation

Medium Term

  • Search index: Build a searchable index across all 1,013 transcripts
    • Full-text search for finding specific topics across all transcribed videos
    • Could use SQLite FTS5 or Elasticsearch
  • Video-to-transcript mapping: Link timestamps back to YouTube video URLs
    • Would allow "jump to 5:30 in this video" from transcript search results
  • Quality validation: Compare transcripts against YouTube auto-captions
    • Spot-check accuracy, especially for technical terms (NativeScript, Angular, etc.)

Backlog

  • Batch resume improvements: Track per-file status in JSON (not just file existence)
  • Parallel processing: Process multiple files concurrently (separate GPU streams for ASR)
  • Whisper comparison: When PyTorch Blackwell kernels mature, benchmark Whisper vs Parakeet
  • pyannote 4.0+ migration: When torchcodec gets ARM64 wheels, upgrade from 3.1.1
  • CMMC v3.0 training data: Potential use of transcripts for fine-tuning data pipeline

Performance Reference

Mode Speed Full Run Est.
Plain text ~110x realtime ~1.4 hours
Timestamps only ~110x realtime ~1.5 hours
Diarized (CPU) ~3.2x realtime ~53 hours
Diarized (GPU, future) ~10x realtime (est.) ~17 hours (est.)