Whisper_speaker_diarization

Whisper speaker diarization on T4 using faster-whisper (large-v3, fp16) and pyannote.audio 3.1. Outputs CSV + SRT.

UI Highlights

• Two-pane brutalist workspace — 4-step stepper (Source · Configure · Process · Review) on the left, tabbed results dock (Transcript · Audio · Downloads · Diagnostics) on the right.
• Stage-aware progress — single PROCESS button auto-chains YouTube download → convert → transcribe → diarize → align → format, with a live monospace log card (first feedback under 2s).
• Interactive transcript — per-speaker color chips with a deterministic palette, click-row-to-seek audio playback, and editable speaker rename that rewrites CSV + SRT in place without re-running the pipeline.
• Inline error card — known failure modes (missing HF_TOKEN, GPU OOM, YouTube auth) render an actionable remediation card next to the progress log.

Required Setup

This Space depends on the gated pyannote/speaker-diarization-3.1 model. Before the app can run you must:

1. Visit https://huggingface.co/pyannote/speaker-diarization-3.1 and accept the user agreement (also accept the pyannote/segmentation-3.0 agreement linked there).
2. Create a Hugging Face access token at https://huggingface.co/settings/tokens with read scope.
3. Add the token as a private Space secret named HF_TOKEN. Locally, export HF_TOKEN=... or place it in .env.

Without HF_TOKEN and the accepted agreements, diarization will fail with a clear runtime error.

Persistent Cache

Model weights (~3 GB Whisper + ~2 GB pyannote) are cached to MODEL_CACHE_DIR. Resolution order:

• MODEL_CACHE_DIR env var, if set (recommended: a Space persistent storage path such as /data/models)
• Otherwise falls back to output/cache/models inside the Space (ephemeral; re-downloaded on cold start)

When the cache dir is resolved the app sets HF_HOME, HUGGINGFACE_HUB_CACHE, and TORCH_HOME to the same location so all framework caches share it.

YouTube Downloads On Hugging Face Spaces

Some YouTube videos require authenticated cookies before yt-dlp can download them. Browser-cookie auth and YouTube username/password login do not work in Hugging Face Spaces because the container does not have your local browser profile and YouTube blocks password login for yt-dlp.

To support those videos, export fresh YouTube cookies in Netscape cookies.txt format and add them as a private Space secret named YTDLP_COOKIES_CONTENT, then restart the Space. Locally, you can also place the exported file at cookies.txt or set YTDLP_COOKIES_FILE to another path.

Downloaded YouTube videos in output/youtube/ are cleaned up automatically after 7 days (files newer than 1 hour are skipped to avoid race conditions with in-flight jobs).