QA 1,000 voice-agent calls with speech-to-text on Sume

Transcribing 1,000 six-minute call recordings with Sume STT costs about $60 at the public rate of $0.01 per audio minute, and each job reserves for up to 10 minutes. It gives you text and word timings to review a voice agent such as Decagon Voice 3 after the fact.

What you get per call

stt_create takes a public HTTPS audio_url and returns text, language code, words with start and end times, and optional sentence segments. Pass duration_seconds (1 to 600) so the reservation matches the call, and a language_code hint such as en if you know it. Without a duration, Sume reserves one minute.

Sume's STT job returns word timings for every call, so you can flag silences, long gaps before an answer, or the exact second a caller interrupted.

The budget

Using the rate card figure of $0.01 per audio minute:

Limits to plan around

A single job reserves a maximum of 10 minutes, so a longer call must be split first. Audio detach extracts audio from a video as WAV or MP3 and can output 16 kHz mono, which is the speech-to-text shape, for $0.01 per job. Its output is capped at 900 seconds, and a longer track needs a range.

Calls must be reachable on a public HTTPS URL, preferably on the Sume media host. Import private files first.

Sample before you scale

You rarely need to read all 1,000. Transcribe a random 5 percent, score it against a checklist (greeting, verification, resolution, handoff), and only expand if the sample shows problems. Use dry_run to get the cost before each batch, and max_spend_usd to cap it.

Sume's STT 1.0 fixes diarization and audio-event tagging server-side and rejects diarize or tag_audio_events in the request, so you cannot ask for speaker labels. If you need to tell the agent from the caller, keep them on separate tracks in your recording.