Tubescript

TubeScript — Speaker Diarization + Transcription

Local pipeline that ingests a YouTube URL (or batches of them), splits audio by speaker using pyannote.audio 3.1, transcribes each segment with Whisper, and exports speaker-labeled transcripts as SRT / VTT / TXT.

This is the missing layer for multi-speaker editing — the existing /youtube-transcripts skill only pulls YouTube's built-in captions (one block of text, no speaker labels). TubeScript answers "who said what when" so the output can be fed into:

• /channel-breakdown — pull quotes attributed to a specific guest in a podcast appearance
• /viral-clipper — auto-cut clips that contain only one speaker's segments
• /transcript-analyzer — extract decisions/action items per speaker
• General podcast/interview content where guest vs host attribution matters

---

Install location

Path	What
C:\Users\A\Desktop\TubeScript\	Repo root (cloned from github.com/Davenads/TubeScript)
C:\Users\A\Desktop\TubeScript\backend\	FastAPI + pipeline code
C:\Users\A\Desktop\TubeScript\backend\venv\	Python 3.12 venv (NOT 3.14 — pyannote/torch/whisper not 3.14-ready yet)
C:\Users\A\Desktop\TubeScript\backend\venv\Scripts\python.exe	Python interpreter to invoke for everything
C:\Users\A\Desktop\TubeScript\backend\.env	Holds HUGGINGFACE_TOKEN=... (you create this)
C:\Users\A\Desktop\TubeScript\frontend\	Vite/React UI (optional — backend-only flow works too)
C:\Users\A\Desktop\TubeScript\start.bat	One-shot launcher (backend + frontend windows)

Note: README/start.bat both expect the venv at backend\venv\ (NOT repo\.venv\). Stick with that — the launcher script depends on it.

---

One-time setup

1. HuggingFace token (REQUIRED — pyannote models are gated)

1. Sign in at https://huggingface.co

2. Accept terms of use on BOTH model pages (one click each):

   • https://huggingface.co/pyannote/speaker-diarization-3.1
   • https://huggingface.co/pyannote/segmentation-3.0

3. Create a token at https://huggingface.co/settings/tokens (Read scope is enough)

4. Drop it into C:\Users\A\Desktop\TubeScript\backend\.env:

   HUGGINGFACE_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxx
   

Without all three steps, diarization throws gated repo and falls over. Whisper does NOT need a token.

2. FFmpeg on PATH

yt-dlp and whisper both shell out to ffmpeg. Verify with ffmpeg -version. If missing, install via winget install ffmpeg or drop a static build into a PATH dir.

3. First-run model preload (downloads ~3-5 GB)

"C:\Users\A\Desktop\TubeScript\backend\venv\Scripts\python.exe" "C:\Users\A\Desktop\TubeScript\backend\preload_models.py"

Pulls Whisper large (~3 GB) + pyannote diarization-3.1 (~500 MB) into HF cache. Skip and they'll lazy-download on first job — preloading just makes the first transcription not stall for 5 minutes.

4. Frontend deps (only if you want the web UI)

cd C:\Users\A\Desktop\TubeScript\frontend
npm install

---

Running it

Option A — full stack via launcher (recommended)

C:\Users\A\Desktop\TubeScript\start.bat

Opens:

• Backend FastAPI → http://localhost:8001
• Frontend Vite UI → http://localhost:3000

Paste a YouTube URL into the UI, set diarization sensitivity (0.5 default = looser, 0.7+ = more speakers detected), wait, then rename SPEAKER_00 / SPEAKER_01 / ... to real names and export SRT/VTT/TXT.

Option B — backend only (script the API)

"C:\Users\A\Desktop\TubeScript\backend\venv\Scripts\python.exe" "C:\Users\A\Desktop\TubeScript\backend\app.py"

Then POST to the API:

# Single video
curl -X POST http://localhost:8001/api/process \
  -H "Content-Type: application/json" \
  -d '{"url": "https://youtu.be/VIDEO_ID", "diarization_enabled": true, "diarization_sensitivity": 0.5}'

# → returns {"job_id": "...", "status": "queued"}

# Poll status
curl http://localhost:8001/api/status/<job_id>

# Get transcript
curl http://localhost:8001/api/transcript/<job_id>

# Export
curl http://localhost:8001/api/export/<job_id>?format=srt -o out.srt
curl http://localhost:8001/api/export/<job_id>?format=vtt -o out.vtt
curl http://localhost:8001/api/export/<job_id>?format=txt -o out.txt

Batch (whole playlist or recent channel videos): POST /api/batch-process with {"url": "...", "limit": 10}. See BATCH_PROCESSING.md in the repo.

Option C — feed a local audio/video file (not a YouTube URL)

The repo's public API only takes YouTube URLs. To diarize a local file, call the modules directly:

import asyncio
from modules.diarization import perform_diarization
from modules.transcription import transcribe_segments
from modules.assembler import assemble_transcript

async def main():
    diarization = await perform_diarization("C:/path/to/audio.wav", sensitivity=0.5)
    segments = await transcribe_segments("C:/path/to/audio.wav", diarization)
    transcript = assemble_transcript(segments)
    print(transcript)

asyncio.run(main())

Run inside the venv from backend/ so the modules.* imports resolve.

---

Output formats

All three carry speaker labels:

• SRT — [SPEAKER_00] Hello there. inside subtitle blocks. Drop into Premiere/DaVinci.
• VTT — same, WebVTT cue format. Drop into HTML5 <track>.
• TXT — SPEAKER_00 [00:01:23]: Hello there. blocks. Best for feeding into LLMs / /transcript-analyzer.

Rename SPEAKER_00 → real name via POST /api/rename/<job_id> BEFORE exporting (mapping persists for that job).

---

Hardware

• GPU strongly recommended. Repo says NVIDIA RTX 4070 Super or better. CPU works but a 30-min podcast takes ~hours instead of minutes.
• VRAM: ~6 GB for Whisper large + pyannote loaded simultaneously.
• Use whisper.load_model("medium") instead of large (edit backend/modules/transcription.py) if VRAM-constrained — quality tradeoff is small for English.

---

Tuning

Knob	Where	Effect
diarization_sensitivity	API param 0.0–1.0	Higher → detects more speakers (and more false positives)
Whisper model size	transcription.py (tiny/base/small/medium/large)	Bigger = slower but more accurate. large is the default.
Min speaker duration	pyannote pipeline params	Filters out quick interjections labeled as separate speakers

For interviews with 2 people: leave defaults at 0.5. For panel discussions (4+ speakers): bump to 0.65–0.75. For monologue with occasional guest: drop to 0.3 to suppress over-segmentation.

---

KNOWN ISSUES from initial install (2026-05-07)

These were observed on this machine after running pip install -r requirements.txt. Fix BEFORE first run.

Issue 1: torch installed CPU-only (no CUDA acceleration)

pip install torch>=2.0.0 from the loose requirement pulls CPU wheels by default on Windows. Verify:

"C:\Users\A\Desktop\TubeScript\backend\venv\Scripts\python.exe" -c "import torch; print(torch.__version__, torch.cuda.is_available())"
# If output ends with "False" → CPU only, transcription will be brutally slow

Fix — reinstall torch with the CUDA 12.1 wheels (matches the 50-series + 4070 super):

"C:\Users\A\Desktop\TubeScript\backend\venv\Scripts\python.exe" -m pip uninstall -y torch torchaudio torchvision torchcodec
"C:\Users\A\Desktop\TubeScript\backend\venv\Scripts\python.exe" -m pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu121

Issue 2: pyannote.audio 4.0.4 installed (not 3.x as repo expects)

The requirements pin pyannote.audio>=2.1.1 so pip resolves to current 4.0.x. Repo code references pyannote/speaker-diarization-3.1 model which still loads, BUT 4.0 introduced the torchcodec dependency that fails to load on Windows without ffmpeg DLLs visible to the linker. You'll see a wall of Could not find libtorchcodec_coreN.dll warnings on import.

Workarounds (pick one):

• Pin to last 3.x: pip install "pyannote.audio>=3.1,<4.0"
• Or install the FULL ffmpeg "shared" build (DLLs, not just the .exe) and add the bin dir to PATH. The "essentials" winget build is exe-only.
• Or feed audio as preloaded {"waveform": tensor, "sample_rate": int} dicts (the warning text suggests this).

The warnings are non-fatal IF you only use the YouTube path (yt-dlp downloads to wav, then pyannote reads the wav via its non-torchcodec fallback). They become fatal if you call Pipeline(...) on a video file directly.

Issue 3: Python 3.14 incompatible

Global Python on this box is 3.14, but pyannote.audio and whisper have no 3.14 wheels yet. Install was forced to Python 3.12 via py -3.12 -m venv .... If you ever recreate the venv, use py -3.12, NOT plain python.

---

Common failure modes

Symptom	Fix
gated repo / 401 from pyannote	Forgot to accept terms on the two HF model pages — see Setup §1
HUGGINGFACE_TOKEN not found	.env missing or backend started outside backend/ directory (dotenv looks at cwd)
CUDA not available, using CPU warning	See KNOWN ISSUE 1 above — reinstall torch with cu121 wheels
Could not load libtorchcodec_core*.dll	See KNOWN ISSUE 2 above — pin pyannote.audio<4.0 OR install full-shared ffmpeg
ffmpeg: command not found (during yt-dlp)	Install ffmpeg, restart shell
Frontend ports clash	start.bat says backend=8001, frontend=3000 in the echo but README says 8000/5173. Check start.bat and frontend/vite.config.js for actual ports — start.bat is authoritative
All speakers labeled SPEAKER_00	Sensitivity too low, or audio is genuinely single-speaker, or the two voices are too similar (same gender + similar pitch + bad mic). Try sensitivity 0.7+

---

Integration cheatsheet

YouTube interview URL
    │
    ▼
[ TubeScript ]  ── speaker-labeled SRT/TXT
    │
    ├──► /transcript-analyzer  → "what did the guest claim about X"
    ├──► /viral-clipper        → cut only guest's punchy lines
    └──► /channel-breakdown    → quote-attribution in essay videos

The whole point: anywhere the existing transcript pipeline drops "wall of text with no attribution," TubeScript replaces it with "labeled dialogue you can route per-speaker."