Model: whisper-small (CPU) | Cost: $0 per transcript | Date: 2026-11 | Status: working
YouTube's "Sign in to confirm you're not a bot" wall ignored a fresh cookie file, a VPN exit, and the server's own IP. The flag isn't tied to the network. It's tied to the Google account.
I run a transcript proxy for a handful of content sites. The chain is cache → YouTube's own captions → local Whisper → a paid API. The goal was zero paid credits. The wall nearly killed the free path, and I spent three days misdiagnosing it.
The wrong claim
An old note in the docs said yt-dlp can't download YouTube from a datacenter. That was wrong. yt-dlp downloaded fine — until the wall appeared. Downloads work in windows, then "Sign in to confirm" shows up.
The note had recorded the symptom and blamed the tool. The tool was never the problem; the account standing was. It's a small distinction that cost me a day: I nearly built around yt-dlp entirely before testing whether it actually failed.
Three layers, three failure modes
# transcript chain — first non-empty wins
cache -> captions -> local-whisper -> supadata | Layer | Cost | What actually happened |
| YouTube captions via a Cloudflare worker | free, ~0.5 s | walled ~50% of calls (LOGIN_REQUIRED), retry up to 8× |
| Whisper-small on CPU, in a container | free, ~1.5× media length | worked, but OOM'd at a 1.5 GB limit |
| Supadata (paid fallback) | per call | 429 plan-exhausted; ~40 videos burned attempts |
The caption layer runs on Cloudflare egress IPs and fetches YouTube's own captionTracks. It's fast — ~0.5 s when it works — but those IPs are walled intermittently, roughly half of calls. The provider retries up to 8 times and bails on no-tracks immediately. Videos without any captions fall through to Whisper.
The Whisper layer was the surprise. The container had base patched to small — a change that's lost whenever the container is recreated. small OOM'd at a 1.5 GB memory limit; bumping the container to 4 GB fixed it. Throughput is ~1.5× media length on CPU, and the sync API tops out at 600 s, so anything over ~10 minutes goes through an exec path instead.
End-to-end it worked: three real videos came back at 3.2k, 6.0k and 9.0k chars of usable transcript. Zero paid credits.
Two gotchas surfaced while scaling it. Parallel Whisper jobs are ~2× slower each, so the queue is serialized behind a lock file. And the paid fallback is a trap when it's throttled: during the caption outages ~40 videos burned transcript attempts on 429s before anyone looked, and some may now be capped at 5 attempts each — meaning the fallback silently consumed the retry budget of videos it never actually transcribed.
The discovery
I tested it like a network problem first, because every prior bot-wall I'd hit was IP-scoped. The free path worked in bursts, then stopped. I rotated the server's IP. Then a VPN exit. Both hit the wall "after a few downloads".
Then I installed fresh cookies exported from the account owner's browser — a 32-line cookie file. Re-tested immediately: the wall persisted on the direct connection and on the VPN exit.
That's the moment the model snapped into place. If fresh cookies on a brand-new network still trip the wall, the wall isn't keyed to the IP. It's keyed to the Google account that owns the cookies.
The numbers back it up. ~30 rapid calls tripped the wall even with cookies installed. The whole day was ~35 test calls before the account cooled off. A flagged account is on cooldown, not banned — back off and it recovers, but keep hammering and the cooldown resets.
The fix wasn't a better proxy. It was pacing:
# after any yt-dlp bot-wall error, skip local Whisper for 15 min
# so a flagged account isn't hammered the moment it trips The scheduler polls channels every 30 minutes, which is naturally spaced enough to stay under the wall. The 15-minute circuit breaker is the important part: the moment a walled account trips, every subsequent job for the next 15 minutes is skipped rather than adding another failed call to the pile.
Bounded daily volume is the whole game. At 30-minute polls with limited backfill, the scheduler tops out at a few hundred calls a day across every channel — comfortably under the trip threshold, which took ~30 rapid calls to trigger. A handful of unrushed calls never flags an account; a burst of them flags it for a day.
What I'd change
Use a dedicated throwaway Google account for yt-dlp cookies, so a transcript outage never flags the account that owns the channels. And treat the wall as a rate limit with a long cooldown, not an IP block. The wrong mental model cost two days of testing proxies that were never going to help — the wall would have followed me to any IP I bought.