STT with no punctuation: Sume still cuts sentence segments on silence

Some speech never gets punctuation: a rambling voice memo, a meeting with crosstalk, a language where the recognizer returns few full stops. Sume's STT API can still hand back sentence-sized pieces. Per the schema, segmentation: {mode: "sentence"} groups words into sentences on terminal punctuation, and splits unpunctuated runs on silence.

The pieces you get

Word timings are always returned, so there is no flag for them. Add segmentation and you also receive sentence segments derived from those word times. boundary_lead_ms (0 to 500, default 70) is how many milliseconds past a sentence's last word the segment runs before the next begins, the same rule sume/tts-1.0 uses.

Fails closed

If the provider returns no timed words, the segmentation request does not guess. Sume returns a typed error. That keeps you from building captions on invented timings.

A request

The audio must be a public HTTPS URL, up to 600 seconds. Sending duration_seconds lets Sume reserve the right amount; without it, Sume reserves one minute. Price is $0.01 per audio minute.

import os

import requests

resp = requests.post(
    "https://api.sume.com/v1/stt-1.0/transcribe",
    headers={
        "Authorization": "Bearer " + os.environ["SUME_API_KEY"],
        "Idempotency-Key": "memo-001",
    },
    json={
        "audio_url": "https://media.sume.com/example/memo.wav",
        "duration_seconds": 95,
        "segmentation": {"mode": "sentence", "boundary_lead_ms": 70},
    },
    timeout=30,
)
print(resp.status_code, resp.json()["data"]["job"]["id"])

Tuning

Raise boundary_lead_ms toward 200 if cuts clip the ends of words; lower it to 0 for tight cuts. A 95-second memo costs about 2 cents at $0.01 per audio minute. Check a few segments by ear. Silence-based splits follow pauses, not grammar, so a long breathless sentence stays one segment.