2026-10-05

TTS sentence slices: mp3 gives timings only, wav gives audio_urls

Sume TTS segmentation returns sentence timings for any container, but slice audio_urls only with wav or raw. Request shape, the 70 ms rule and when to pick wav.

4 min readSume

All posts

Sume TTS can cut one voiceover into sentence pieces for you, which saves one job per sentence. There is a catch in the request schema: segmentation.emit_audio defaults to true, but slices are returned as audio only when output_format.container is wav or raw. With mp3, the job returns timings and no per-segment audio_url.

What you must send

timestamps: {words: true}: segmentation requires word timings.
segmentation: {mode: "sentence"}: the only mode in v1.
output_format: {container: "wav", sample_rate: 44100, encoding: "pcm_s16le"} for audio slices.

What comes back

Segments are gapless: segment[i].end equals segment[i+1].start. A sentence ends 70 ms after its last word by default (boundary_lead_ms, 0 to 500), and the next segment takes the pause. With wav or raw each segment carries a sample-exact audio_url.

Pick the container by the next step

Container choice for segmented TTS
Next step	Container	Why
Per-sentence clips for a timeline	wav	Slices get audio_url
One narration file for a podcast feed	mp3	Small file, timings still returned
Join takes without gaps later	wav	Sample-domain concat
Phone system	wav with pcm_mulaw, 8000 Hz	Telephony format

A request body

Send it to POST /v1/tts-1.0/generate with your voice selector and an Idempotency-Key.

import json

body = {
    "transcript": "First point. Second point. Third point.",
    "voice": {"mode": "id", "id": "YOUR_VOICE_ID"},
    "timestamps": {"words": True},
    "segmentation": {"mode": "sentence", "boundary_lead_ms": 70},
    "output_format": {
        "container": "wav",
        "sample_rate": 44100,
        "encoding": "pcm_s16le",
    },
}
print(json.dumps(body, indent=2))

Cost

Segmenting does not add a line to the bill; you pay for characters. Three short sentences of 40 characters are 120 characters, 1 cent, the minimum. A 20,000-character script is $0.95 whether or not you segment it.