PII redaction on 40 hours of calls: AssemblyAI $14.40 vs Sume STT

For 40 hours of call recordings with PII redacted, AssemblyAI's published prices stack to $14.40 on Universal-3.5 Pro with both redaction add-ons and speaker diarization. Sume STT would cost $24.00 for the same 40 hours of audio, and Sume does not list an automatic PII redaction option, so you would have to find the sensitive spans yourself.

Prices below were read on each vendor's own page on 2026-10-09 and on Sume's public catalog. They change, so check before you commit a budget.

What AssemblyAI charges per hour

AssemblyAI prices asynchronous transcription as a base model plus add-ons that stack additively. The base for Universal-3.5 Pro is $0.21 per hour. PII audio redaction adds $0.05 per hour and PII text redaction adds $0.08 per hour. Standard speaker diarization adds $0.02 per hour.

For a call archive where you need a redacted transcript and a redacted audio file, all four lines apply. That is $0.21 + $0.05 + $0.08 + $0.02 = $0.36 per hour.

What Sume charges for the same audio

Sume STT 1.0 is priced at $0.01 per audio minute in the public catalog. Forty hours is 2,400 minutes, so 2,400 x $0.01 = $24.00. That price is for transcription with word timings only. The request body has no PII field, and the diarize setting is fixed server-side rather than something a caller can switch on.

A single STT job is capped at 10 minutes of reserved audio (duration_seconds up to 600), so 40 hours of calls is at least 240 jobs if each call is split into ten-minute slices. Calls shorter than ten minutes map one to one.

Where Sume can still help with redaction

If your redaction rule is simple, such as removing a known customer name, the word timings Sume returns let you cut those spans out of the audio yourself. Timeline audio can split a hosted audio file into as many as 20 ranges per job at a flat $0.01, and concat the pieces back together, which is how the redaction pattern in the linked post works. That is a manual, rule-based process. It does not detect PII for you.

So the honest summary is that AssemblyAI is cheaper per hour and does the detection. Sume is worth considering only if you already run your media pipeline through Sume and your sensitive strings are known in advance.

• Need automatic detection of card numbers, addresses or names you cannot list: use a vendor with a redaction feature.
• Need known-string removal inside an existing Sume pipeline: STT words, then timeline audio split and concat.
• Always have a person spot-check a sample of redacted output; neither route is a compliance guarantee.

A worked decision rule

Start from what your rule says about the data, not from the price. If a regulation or contract requires that card numbers, health details or addresses never reach a reviewer, detection has to be automatic, because a person cannot reliably hear every case in 40 hours of audio. That points to a vendor whose product detects and redacts, and the $14.40 line is the relevant number.

If the sensitive items are a short, known list, for example the names of the 30 customers on the calls, the manual path is realistic. You transcribe once, search the word list for those names, and cut the matching ranges. The cost is the STT charge plus a few one-cent timeline jobs, and the engineering time to write the search.

Whichever route you use, keep the original recordings under your own access controls until the redacted set has been reviewed. Redacting a copy is safer than editing the only file you have.