YouTube transcripts via yt-dlp audio + Whisper transcription

Research Question

Can we bypass YouTube's IP-based transcript block by downloading the audio track with yt-dlp (a different endpoint from the transcript API) and then transcribing it with OpenAI Whisper?

Findings

Executive Summary

The hypothesis that yt-dlp audio CDN downloads are less restricted than the transcript API on GitHub Actions IPs is false: audio and video both traverse the same YouTube CDN infrastructure and are subject to the same cloud IP block. This means the proposed fallback chain (caption fetch fails → audio download → local Whisper) fails at the second step from any GitHub Actions runner. If the audio download barrier is cleared — most likely via browser cookies stored as a GitHub secret — faster-whisper with int8 quantization on the small model provides a practical transcription path (~7–15 min per 60-min talk, free). Option B (Whisper API) is blocked by the credential gate: OPENAI_API_KEY is not an approved credential.

Key Findings

1. [high] YouTube delivers audio and video streams via the same CDN infrastructure; cloud IP restrictions that block yt-dlp video downloads also block yt-dlp audio-only downloads from GitHub Actions runners running on AWS IP ranges.
2. [high] As of 2024–2025, yt-dlp downloads from cloud/datacenter IPs (including GitHub Actions) are actively blocked by YouTube, producing HTTP 403 errors or "Sign in to confirm you're not a bot" challenges, regardless of whether the requested format is video or audio.
3. [high] The only currently identified workaround that avoids paid third-party infrastructure is passing browser-exported cookies to yt-dlp via --cookies, stored as a base64-encoded GitHub Actions secret — but this requires periodic manual maintenance as cookies expire within weeks to months.
4. [high] Using vanilla openai-whisper with the small model on a CPU-only GitHub Actions runner, transcribing a 60-minute audio file takes approximately 72–120 minutes — not the "3–5 minutes" estimated in the original item, which appears to assume GPU availability.
5. [high] faster-whisper with int8 quantization delivers 4–7x speedup over vanilla openai-whisper on CPU, reducing the same 60-min audio to ~7–15 minutes of runner time, making the approach economically viable on the free Actions tier.
6. [medium] The Whisper small model achieves 3.2–3.4% WER on clean English speech; medium achieves 2.7–2.9% WER — a 0.5–0.7 percentage point improvement that widens for noisy or accented audio but is modest for well-recorded English talks.
7. [high] The small model requires ~461 MB download; medium requires ~1.5 GB. Both should be cached via actions/cache to avoid re-downloading on every workflow run.
8. [high] Option B (OpenAI Whisper API at whisper-1, $0.006/min) requires OPENAI_API_KEY, which is not in the AGENTS.md approved credentials table; it cannot be implemented without explicit owner approval.
9. [medium] The OpenAI Whisper API whisper-1 model is equivalent in quality to large-v2 locally, delivering higher accuracy than any local model size runnable on GitHub Actions CPU within a reasonable time budget.
10. [medium] Third-party transcript APIs (the companion backlog item 2026-02-28-transcript-via-third-party-apis.md) sidestep the cloud IP block entirely and may offer a better cost/complexity tradeoff than maintaining YouTube cookies in a GitHub secret.

Assumptions

• [assumption] GitHub Actions runners use AWS IP ranges. Justification: GitHub-hosted runners are known to run on Azure and AWS infrastructure; both are in cloud IP ranges that YouTube blocks. This assumption is consistent with the prior research item's documented block.
• [assumption] The cookies workaround would require the owner to use a personal Google account. Justification: The research corpus does not specify a dedicated YouTube account. Using a personal account exposes personal cookies to GitHub Secrets, which is a security consideration.

Analysis

The evidence converges on a clear conclusion: the hypothesis is false, and the proposed architecture cannot work without first solving the same cloud IP block problem that the transcript API approach already fails on.

The yt-dlp + Whisper approach has genuine merit as a Tier 2 fallback if audio download succeeds — the faster-whisper int8 optimisation resolves the runtime concern entirely, and the small model accuracy is adequate for research transcription. The approach becomes viable under one of: (a) the cookies workaround is accepted by the owner with its maintenance requirements; (b) a third-party proxy is interposed; or (c) YouTube's blocking policy changes.

The original item's runtime estimate ("~3–5 min for a 60-min talk on a free GitHub runner") deserves correction: it appears to have assumed GPU availability. The ubuntu-latest free runner is CPU-only, and vanilla whisper-small takes 72–120 minutes on that hardware. faster-whisper resolves this but was not mentioned in the original design. Any implementation should use faster-whisper, not openai-whisper.

Option B was correctly identified as the simpler path, but it introduces an unapproved credential. The cost ($0.36/hour) is low enough to be acceptable for a personal research workflow, but the credential gate must be addressed first.

Risks, Gaps, and Uncertainties

• Empirical testing not performed. The blocking of yt-dlp from GitHub Actions is well-documented in community reports but not tested in this specific repository's workflow. A 5-minute test step in fetch-transcript.yml would confirm the block definitively.
• Cookies lifespan is variable. There is no community consensus on the exact expiry window; some report weeks, others months. The operational burden of the cookies approach is therefore uncertain.
• faster-whisper accuracy at int8. The benchmark comparison between float32 and int8 Whisper small shows negligible WER difference in most tests, but this is not universally verified for domain-specific technical vocabulary in research talks.
• Sources marked [ ]: The three primary sources in the item's Sources section (yt-dlp GitHub, openai/whisper GitHub, OpenAI platform docs) were not directly read; they are represented by secondary evidence. Their content is well-characterised in the literature.

Open Questions

• Should the owner approve OPENAI_API_KEY for the Whisper API path? At $0.006/min, a year of weekly 60-min talk transcriptions would cost ~$18.72 — a low barrier for the accuracy gain.
• Does the third-party API research item (2026-02-28-transcript-via-third-party-apis.md) yield a cleaner solution that avoids cookies maintenance entirely?
• What is the empirical yt-dlp exit code from this repository's GitHub Actions runner? A single test step would close the uncertainty about whether this runner's IPs are actually blocked.

Open Questions

• Are yt-dlp audio CDN endpoints blocked from GitHub Actions AWS IP ranges?
• What is accuracy difference between small and medium Whisper models for academic talk audio?
• Is the runtime acceptable on a free GitHub runner for talks > 60 minutes?

sources

• [ ] yt-dlp docs
• [ ] OpenAI Whisper (local)
• [ ] OpenAI Whisper API