Trim the silent lead-in: first spoken word from STT, then video trim

How do you remove the silence before someone starts talking? Ask speech-to-text when the first word begins, then cut the video from just before that time. Sume's STT returns words[] with start and end seconds, and POST /v1/video-trim cuts [start, end) into a new MP4.

Fast first partials are a selling point for live models: Microsoft says MAI-Transcribe-2-Streaming produces first text just over 100 ms after audio arrives (source, read 2026-10-04). For a finished clip you do not need that, only an accurate first timestamp.

Get the first word

Video inspect with transcribe: true runs Sume STT on the clip's audio and returns transcript with text and words[]. The transcript adds the STT rate, documented as $0.01 per audio minute, to the inspect reservation. Word items can include a type of word or spacing when the provider supplies one, so skip any spacing token when you look for the first word.

Cut with a small pad

Start the cut a little before the word so the breath survives. Use start plus exactly one of end or duration; sending both returns video_trim_range_conflict. The default precision: exact is a frame-accurate re-encode. precision: keyframe is a stream copy whose cut may begin a GOP early, so re-base against actual_start_seconds in the result.

A trim costs $0.02 per job per the video trim guide; confirm live in GET /v1/catalog.

import os, requests

first_word_start = 2.31      # words[0].start from the STT result
clip_duration = 24.0         # probe.duration_seconds from video inspect
pad = 0.15
start = max(0.0, first_word_start - pad)
r = requests.post(
    "https://api.sume.com/v1/video-trim",
    headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"],
             "Idempotency-Key": "trim-leadin-001"},
    json={"video_url": "https://media.sume.com/artifacts/artf_demo/talk.mp4",
          "start": round(start, 2), "end": clip_duration,
          "precision": "exact"})
print(r.status_code)

Edge cases

If end runs past the source it clamps and the result warns trim_clamped_to_source. If the clip has no speech at all, the transcript has no first word; handle that case instead of cutting at zero. Output must be at least 0.2 seconds and at most 900 seconds.