OpenAI TTS vs Audra: loudness and long-form limits — Audra

OpenAI's speech endpoint is cheaper per character than Audra — $15/1M on tts-1 versus $25/1M at checkout. This comparison is not about sticker price; it is about what a long-form pipeline pays in chunking, mastering, and request shaping once files exceed a few thousand characters.

The long-form problem

A 45-minute audiobook chapter is roughly 34k characters of script. OpenAI caps each speech request at 4,096 characters, so the chapter becomes nine round trips plus your own concat logic — and a failure on request seven of nine means re-slicing mid-sentence. Audra takes up to 50k characters in one synchronous POST and 500k through the jobs API, so the same chapter is one request and one file.

Loudness: what −16 LUFS saves you

Podcast and audiobook distributors normalize around −16 LUFS. OpenAI returns raw audio; meeting spec means a loudnorm pass through ffmpeg per file, plus the storage and orchestration to run it. Audra bakes −16 LUFS into every MP3 export, so the file that lands in your bucket is the file you publish.

Where each one wins

The 50k sync limit and per-request shaping are documented at /docs; the full feature matrix is at /vs/openai-tts.

Next steps