Cozzo
Find yours

Cozzo-tested

How we’ll test voice AI ourselves

Prices tell you what a voice agent costs. They don’t tell you how fast it answers, how well it hears, or how it sounds. So we’re running our own tests, and publishing exactly how before any result goes up.

Testing in progress — no scores published yet

Nothing on Cozzo is a Cozzo-measured latency, accuracy or naturalness score yet. Latency figures elsewhere on the site come from vendors’ own docs, not from our tests.

Cozzo tip: I haven’t timed anyone yet, so you won’t find a score here. When I do, every number will say how many calls it came from.

What we’ll measure

Time to first audio

How long a caller waits before the agent says anything.

How

Measured from the moment the call is answered to the first audible agent audio, on our caller-side recording.

Reported as

Median and 90th percentile (p90) in milliseconds, plus the number of calls.

Response latency

The pause after you stop talking, before the agent answers. This is the gap that makes a bot feel slow.

How

Measured on every turn of every scripted call: from the end of the caller’s speech to the start of the agent’s reply. Includes telephony and network, because a real caller hears those too.

Reported as

Median and p90 across all turns, with the share of turns slower than 1.5 seconds.

Transcription accuracy

Whether the agent heard you right, especially names, numbers and dates.

How

Word error rate (WER) on a fixed set of 200 recorded utterances: clean speech, phone-quality 8 kHz audio, background noise, a range of accents and a Spanish subset. Transcripts are normalized the same way for every vendor (case, punctuation, number formats) before scoring.

Reported as

Overall WER, WER by condition, and a separate accuracy score for names, phone numbers and dates.

Naturalness

Whether the voice sounds like someone you’d want answering your phone.

How

Blind listening test: the same lines from each vendor’s default and recommended voices, shuffled and unlabeled. At least 20 listeners rate each clip from 1 to 5. A human reference clip and repeated clips catch careless ratings.

Reported as

Mean rating with a 95% confidence interval, never a single “winner” badge without one.

The test protocol

The same test for everyone, so differences come from the product, not the setup.

  1. Fixed scripts

    10 call scenarios a real business gets: booking, rescheduling, opening hours, a caller spelling their name and number, an interruption mid-sentence, a request for a human, a noisy line and a Spanish-speaking caller.

  2. Identical caller audio

    Every scenario uses the same pre-recorded human caller audio, played into the call, so every platform hears exactly the same input.

  3. Same job for every agent

    One fictional business with the same prompt, knowledge base and booking rules on every platform.

  4. 50 calls per platform

    Spread across weekdays and times of day. Failed or dropped calls are counted and reported, not quietly re-run.

  5. Measured on the caller’s side

    Calls are recorded from the caller’s end and timed with automatic voice-activity detection, then spot-checked by hand.

What we keep the same

  • Same scripts, caller audio, prompt and knowledge base everywhere.
  • Same LLM wherever a platform lets us choose; otherwise the platform’s default, named in the results.
  • Default or vendor-recommended voice and speech-to-text settings, documented per platform.
  • One US carrier and number type for phone tests, calls placed from the same US East region. Web calls are tested and reported separately from phone calls.
  • Self-serve accounts at list price, the same way you’d sign up. No special endpoints, no vendor-tuned demos.

How results will be reported

  • Each tested vendor gets a results card: latency, accuracy and naturalness numbers, how many calls or clips they come from, the test dates and the exact configuration.
  • Raw per-call measurements will be downloadable, so anyone can check the math.
  • Vendors can review their configuration before results go live, to catch a setup mistake. They can’t edit, delay or veto results.
  • Scores are dated and re-run at least every quarter, and sooner when a vendor ships a major model change.
  • Sponsors and advertisers get no say. Tested scores won’t change finder rankings until they’re published here, and any weight they get will be listed on the methodology page.

First in line

The most-compared products on Cozzo. None has been tested yet.

Platforms

Speech APIs

Questions

Why no scores yet?

Because we haven’t finished measuring. Cozzo doesn’t publish benchmarks it didn’t run itself, and we don’t reuse vendor-published latency claims as test results.

Can a vendor pay to be tested first or scored higher?

No. Testing order follows what visitors compare most. Paying for a sponsor placement has no effect on testing or results.

Will the protocol change?

Possibly, before the first results. If it does, the change and the reason will be logged on this page with a date.

How do I suggest a scenario or a vendor?

Email partners@cozzo.io. Vendors who want to be included can write to the same address.

In the meantime

Compare what’s verifiable today: prices from vendors’ own pages in the cost calculator, vendor-published model specs on voice models, and a shortlist from the finder. Price moves are logged on pricing changes, and the rest of how we work is on the methodology page.