OpenAI charges by the minute, so speed up your audio

Original Article ↗ Hacker News Discussion ↗

Core trick: speeding audio to cut cost/time

  • Original post describes using ffmpeg to speed a 40‑minute talk up 2–3× to fit OpenAI’s 25‑minute upload cap, reducing cost and latency while still getting usable transcripts/summaries.
  • Several commenters report similar discoveries (e.g., 2× for social-media reels) and note it feels “obvious” once you think in terms of model time vs. wall‑clock time.
  • Some point out this is conceptually similar to lowering sample rate or downsampling intermediate encoder layers in Whisper to gain throughput.

Alternatives, pricing, and business angles

  • Multiple people suggest bypassing OpenAI’s transcription API entirely:
    • Run Whisper (or faster‑whisper/whisper.cpp) locally, especially on Apple Silicon.
    • Use cheaper hosted Whisper from Groq, Cloudflare Workers AI, DeepInfra, etc., citing ~10× lower prices.
    • Use other LLMs with audio support (Gemini 2.x, Phi multimodal) or specialized ASR services.
  • Some are already selling “speech is cheap”‑style APIs, arguing you must add value (classification, diarization, UI) beyond raw transcription.

Accuracy, limits, and evaluation

  • People question accuracy at 2–4× speed, asking for word error rate or diff‑based comparisons; others argue what matters is summary fidelity, not verbatim text.
  • Suggestions include:
    • LLM‑based evaluation of whether key themes persist across different speeds.
    • Measuring variance by running the same audio multiple times.
  • An OpenAI engineer confirms 2–3× still works “reasonably well” but with probably measurable accuracy loss that grows with speed.

Local vs. cloud, privacy, and efficiency

  • Strong thread arguing that local Whisper is “good enough,” essentially free, and avoids sending personal interests or sensitive data to OpenAI.
  • Others counter that newer proprietary models (e.g., gpt‑4o‑transcribe) can be faster or better, but can’t be run locally.

Preprocessing tricks and tooling

  • Multiple ffmpeg recipes shared to:
    • Remove silence (and thus cost/time) before transcription.
    • Normalize audio to reduce hallucinations.
  • Many tips on grabbing and using YouTube transcripts (yt‑dlp, unofficial APIs), and on playback‑speed extensions (up to 4–10×).

Meta: speed vs. understanding

  • Substantial side‑discussion:
    • Some argue summaries and 2–3× playback are “contentmaxing” but degrade depth of thought.
    • Others say speeding content just matches their natural processing rate, and depth comes from intentional re‑watching and reflection.