YouTube Audio Transcript API

Speech-to-text for YouTube audio.

Transcribes YouTube audio with AI speech-to-text when the video has no published captions (or when you want speech-to-text regardless). Audio with no speech answers 422 NO_SPEECH at 0 credits, remembered for 24 hours. Separate from /transcript — that endpoint only returns YouTube's caption tracks. Pricing is duration-based and honest: creditsUsed = ceil(durationSeconds / 60) × 2 (badge: 2 credits/min of audio). Success includes estimatedCredits (same number) so maxCredits is usable without doing the arithmetic. Pass maxCredits to refuse expensive jobs before any STT runs (400 cost_exceeds_max, 0 credits). Measured end-to-end: ~20 min of audio ≈ 12s, ~82 min ≈ 49s. Audio is re-encoded to 16 kHz mono 32 kbps before upload so podcast-length jobs stay under the ~25 MB ceiling. Sync path is capped at 90 minutes so the request finishes inside the 110s request deadline; longer videos return 400 duration_too_long with estimatedCredits before any STT spend. An audio download miss is 502 audio_extract_failed with a split reason: blocked_rate (403/429, retryable, retryAfterSeconds=300), extract_timeout (retryable, 60s), extractor_unavailable (signed-URL / extractor miss, retryable, 120s), no_audio_format / restricted / live_unavailable (this video, retryable false). estimatedCreditsIfRetried is only present when retryable is true. A permanent miss with no captions sets obtainable:false and says no transcript is obtainable — it does not leave suggestedEndpoint:null as an open loop. Download budget is 20s (a doomed fetch used to sit for ~67s). blocked_rate opens a service-wide circuit (cacheScope=service, blockedRateConsecutive) so further downloads wait out the cooldown instead of deepening a YouTube block. When a published caption track exists, suggestedEndpoint=/v1/youtube/transcript — suggest, never substitute STT with captions. Response always includes source: "asr", asrProvider, languageIsDetected, numeric segments[{text,startMs,endMs}], text, durationSeconds, creditsUsed, estimatedCredits. Cache hits still bill — the cache is our margin.

Docs: https://captapi.com/apis/youtube-audio-transcript