A 5-minute voice-over: 4.8 MB as 128 kbps MP3, 26.5 MB as WAV

A five-minute voice-over is 4.8 MB as a default Sume TTS MP3 (128 kbps) and 26.46 MB as 16-bit mono WAV at 44.1 kHz, or 28.8 MB at 48 kHz. The WAV sizes assume one channel and 16-bit samples, which is my assumption; stereo doubles them. Choose MP3 when the file goes straight to a listener, and WAV or raw when it goes on to another processing step such as cutting, joining or sentence-level timing.

The arithmetic

MP3 at 128 kbps is 16,000 bytes a second, so 300 seconds is 4,800,000 bytes. Uncompressed audio is sample rate times bytes per sample times channels: 44,100 x 2 x 1 = 88,200 bytes a second, which is 26,460,000 bytes for 300 seconds. At 48,000 Hz the rate is 96,000 bytes a second and 28,800,000 bytes.

Which step needs which

The TTS docs tie sentence timing to the output: emit_audio for sentence segmentation needs wav or raw output. Cutting and joining audio with Timeline audio is cleaner on WAV, because MP3 frames add a little padding at every edge. For a mux onto video, either works. For an upload to a platform or a download link, MP3 is a fifth of the size.

Speech to text accepts an audio_url with its own size limit, so check the OpenAPI reference before you send long WAV files to transcription; the MP3 is a fifth of the size.

A default that holds up

Keep the master as WAV for cutting, create an MP3 for delivery, and measure the real file before you assume a size. Both cost the same TTS price; the format does not change the $0.0475 per 1,000 characters.

Rates other than 44.1 kHz

Output sample rates run from 8,000 to 48,000 Hz. A lower rate shrinks WAV in direct proportion: 8,000 Hz mono 16-bit is 16,000 bytes a second, 4.8 MB for five minutes, the same as the default MP3, but with the quality of a telephone line. Speech models usually want 16,000 Hz, so a 16 kHz mono WAV is a reasonable working copy at 9.6 MB. The size of the file never changes the speech price.