Methodology
How we verify facts, how we compute real production costs, and how we make money. This page is the contract behind every number on the site.
Every fact links to a primary source
Pricing pages, docs, changelogs. A fact without a source URL and a last-verified date does not render publicly. Where a number is a vendor claim (latency figures, for example), it is labeled as a claim on the page until we can measure it independently. Facts still flagged as sample data never render in production at all; you see "verification pending" instead.
Re-verified on a weekly cycle
Every number carries the date we last checked it, and the "What changed recently" section on each comparison is generated from the fact history, not written by hand. Stale facts get flagged for review before they get republished.
We earn either way, so verdicts stay honest
vsref may earn a commission when you start a trial through our links, and every such link is marked and disclosed on the page it appears on. Programs exist on both sides of most comparisons, so no verdict is worth more to us than its opposite. As a structural check, no single platform should win more than 45% of its matchup verdicts site-wide; the live numbers are published below, including when they breach that line.
True cost, not sticker price
Advertised per-minute rates rarely survive contact with production. Our cost tiers model all-in monthly spend (platform fee + per-minute + telephony) at 1K, 10K, and 100K minutes per month, using the measured real cost range where we have one.
One unit per category, converted in code
Vendors price the same work in different units: per character, per credit, per second, or a subscription with a quota. We normalize every rate to one canonical unit per category with deterministic math, never estimates. For text-to-speech that unit is dollars per 1M characters and dollars per audio minute, assuming ~950 characters of English text per minute of finished audio; subscription plans resolve to plan fee divided by included quota, plus published overage. For speech-to-text the unit is dollars per audio minute (and per 1,000 minutes), batch and streaming kept separate: per-hour rates divide by 60, per-second rates multiply by 60, and toggling an add-on like diarization or PII redaction adds its published per-minute rate, or renders as not computable when the vendor prices it but publishes no rate. Credit systems are converted only when the vendor states the credit-to-unit mapping, and a plan quota we cannot convert honestly renders as not computable rather than a guess.
Who reviews this
We publish under the vsref Editorial name, not a personal byline, and we will not invent one. Comparisons and verdicts are drafted from the verified fact layer by the verdict engine, then published through an editorial review flow: a publish guard checks the generated prose against the same sourced, dated facts the tables show and blocks anything that fails the quality gate. That is what "reviewed by vsref Editorial" means on each page, the process is accountable to a primary source rather than to an opinion. If and when a named human reviewer joins, they are credited here by name; until then vsref carries the editorial responsibility as a publisher.
Verdict balance, live
Win rate per platform across all published use-case verdicts, straight from the same database the pages render from. A platform over the 45% line is a signal to us to widen its matchup coverage, not to soften verdicts.
Tool logos are the property of their respective owners and are shown here for identification only; their use does not imply any affiliation or endorsement.