How the Voice Cost Audit works
A 10-business-day audit of one expensive part of your live voice stack, in one language. I measure what one connected minute costs you today, then compare three ways to lower it: tune what you run, switch to a cheaper API, or self-host an open model. I recommend the cheapest option that passes the quality, latency and reliability bar we agree for your calls.
A 20-minute scoping call
No recordings and no invoices. You tell me roughly how many connected minutes you run a month, the language and call type, which speech-to-text, LLM, text-to-speech, platform and phone vendors you use, and what worries you about latency or quality.
- We pick the scope together. One component, one language, a representative workload, and who on your side may see the data.
- You get a one-page scope. Access, dates, how the options will be compared, and the fee.
- An honest warning up front. If you pay under 6 cents a minute and run under 100,000 minutes a month, I'll tell you before you pay that the savings may be small, so you can decide if it's worth it.
What I ask for, after you accept the scope
Anything missing is written down as an assumption in the report, never quietly estimated. I don't ask for production write access to start the audit.
- Invoices. 30 days of itemised vendor, platform and phone invoices or billing exports, with discounts, taxes, commitments and minimums, as PDF or CSV.
- Minutes. A 30-day export of connected minutes and call counts for the scoped call type and language, with how each vendor bills and rounds.
- Usage. Aggregate speech-to-text, LLM and text-to-speech usage (tokens, characters, concurrency, latency, failures and retries, transfers and hold time), as CSV or a read-only dashboard view where you have one.
- A call sample. A representative sample you choose, with any transcripts or reference outcomes you already have and any consent or use limits. It stays inside your own systems and is never emailed to me.
- Your quality bar. The accuracy or task-success rubric, accent and noise mix, latency limits, fallback rules and anything compliance rules out.
- A technical contact. One named person, and read-only or supervised access.
Who sees what
Call audio and transcripts
They stay in your infrastructure; I don't take a copy. I take out only the aggregate measurements and any de-identified examples you approve for the report. Your recordings and billing data are never shared with an alternative API vendor.
The exports you approve
I'm Poojith, the founder, and I run the audit myself. I see the invoice and usage exports and the sample results, through the access you choose. Your contract rates stay in a private calculation; vendor list prices are checked separately.
If a live API test is needed, you choose the vendor and run it from your own account after you approve it. Only then does that vendor receive test audio, under your terms. A list-price comparison never involves a third party.
Ten business days
The ten business days start once the data and access above are in place.
Day 1: Lock scope and permissions
Confirm the component, language, call population, invoice window, data owner and sample access. Write down the quality, latency and reliability pass criteria before any option is compared.
Day 2: Reconcile the baseline
Tie 30 days of charges and credits to 30 days of connected minutes. Split speech-to-text, LLM, text-to-speech, platform, phone and fixed commitments, and note how each vendor bills: connected time, speech time, characters, tokens or rounded call time.
→ your baseline cost per connected minute, with the formulaDay 3: Choose the sample
Inside your environment, pick calls that cover short and long, noisy and clean, accents, silence and hold, peak concurrency and common failures. The count and mix are agreed before any option is tested.
Day 4: Measure today's path
Run your current component on that sample, or read its approved logs. Record quality, p50 and p95 latency, timeouts, retries, concurrency and the billed units. Every candidate has to meet this.
Day 5: Path A, tune your stack
Price changes such as prompt caching, a different model or voice tier, silence handling, token limits or batching, only where they apply to your stack, at your contract rates. Test each change on held-out calls and count the engineering it needs.
Day 6: Path B, a cheaper API
Map the same workload onto each plausible vendor's current billing units, with platform fees, minimums, concurrency, phone and fallback. List prices are marked apart from your contract. A paper price is a candidate, not a passed test: without a live test from your account, its gates are marked unverified.
Day 7: Path C, self-host open models
Estimate compute, idle capacity, throughput at your peak, storage, upkeep and fallback from the measured workload and a real machine quote. If a sandbox inside your systems is available, test a quantized model on the same sample; if not, its gates are marked unverified and no saving is claimed.
Day 8: Compare the three paths
One table: price source and date, billed units, monthly total, all-in cost per connected minute, one-time and ongoing work, quality, latency and reliability status, and why anything was ruled out, with sensitivity to volume, peak concurrency and discounts.
Day 9: Review together
We go through the evidence and every unknown. I recommend an option that passed the gates, or say plainly that none did, and hand your technical owner the steps, the risks and the fallback.
Day 10: Delivery
The baseline table, the sample definition, the three-path comparison, test results, assumptions, the recommendation and the refund decision. See the synthetic worked audit for the shape of it; it is not a client result.
→ one written recommendation
How the quality gate works
Before any test, we agree the call-level rubric, the acceptable error or task-success margin, the p95 latency target, timeout and retry tolerance, and any hard compliance rules. Every path is tested on the same held-out sample. A cheaper option that fails a hard gate is not the recommendation, and anything that couldn't be measured is shown as unverified, never as a pass.
Price and refund
$1,000 fixed, or ₹75,000 for India-based teams. Full refund if the audit finds possible savings under 20% of your current all-in cost per connected minute, from your last 30 days of invoices. Every option counts toward that 20%, including a vendor API.
If self-hosting wins and you choose a pilot with ILLATE, half the audit fee is credited to it. The pilot is separate, takes about six weeks and has its own acceptance test. You don't need a pilot to use the findings. Full terms: the Voice Cost Audit.
Start with 20 minutes
No data needed for the first call. I'll tell you honestly whether I think your cost per minute can come down.