Skip to content
Trueline

Detection performance

Measured on live calls, not on benchmark files.

Detection scores in this category are ordinarily computed over corpora of audio files. The figures below were produced by placing calls over commercial carriers and scoring them with the production pipeline at its default operating point, which is the condition your traffic arrives in.

Measurement period
5 to 11 September 2026
Operating point
Conservative, the shipped default
Engine
Release 9, production weights
Evidence
Call-level, live carrier paths

86%

Detection rate

Synthetic callers detected on internet and carrier routes, across 22 live calls.

0.0%

False positive rate

Genuine callers flagged on clean lines. On degraded lines the measured rate is 5.0%.

12s

Time to first verdict

Of caller speech before a call is decided. Scoring continues for the rest of the call.

Detection

Detection rate by call path.

Conservative rule · call level

  • Inbound AI voice agents, internet routes100%

    Live agents dialling our number over SIP, including an engine never present in training

    Calls 3 · Detected 3Validated
  • Synthetic voices over a commercial carrier84%

    Nineteen agent and cloned voices carried to our number over a PSTN interconnect

    Calls 19 · Detected 16Validated
  • Calls with a mobile handset leg0%

    Handset and mobile-carrier processing removes part of the signal the model scores

    Calls 9 · Detected 0Under validation
  • Internet and carrier routes, combined19 of 22 · 86%

Two of the three carrier misses raised one confident window against a rule that needs two; the rule sits there because a false alarm reaches an agent. Paths with a mobile handset leg are not priced or contracted against: if a material share of your callers are on mobiles, raise it during scoping and the pilot quantifies it on your routes.

False positives

Genuine callers flagged, by line condition.

Genuine callers · call level

  • Clean lines0.0%

    Genuine call audio at the quality a contact centre would consider normal

    Calls 12 · Flagged 0Validated
  • Degraded lines5.0%

    Recorded genuine speech carrying heavy compression, music and crowd noise, scored through the production pipeline

    Calls 40 · Flagged 2Within tolerance

A flag is advisory and never ends, transfers or blocks a call, so a false positive costs an agent a glance rather than a customer.

Methodology

A number is only worth its method.

01Detection
A call is counted as detected only when the flag is raised while the call is still connected. Post-call identification is not counted.
02Decision rule
Two four-second windows scored above threshold within any thirty-second span. The same rule and threshold ship to every tenant by default.
03Evidence base
Calls placed over commercial carriers and scored by the production pipeline. Benchmark corpora are not used, and results are not re-scored offline after the fact.
04Revision
Figures are re-measured at each engine release and this page is dated. Superseded figures are not carried forward.

Validation

These figures are ours. The ones that matter are yours.

A one-week pilot scores your live traffic and returns the same tables for your own routes, issued under NDA whatever the result.

Request a demo →
  1. 01

    Detection rate on your routes

    Call by call, on your own numbers and your own line quality.

  2. 02

    Every flag and every miss

    Each flag with the rule that fired and the second it fired on; each miss with the conditions it arrived under.

  3. 03

    The operating point we recommend

    False positives by line condition, and what moving the rule in either direction would cost you.

Notes

  1. 1.Detection figures cover 22 live calls on internet and carrier routes and 9 calls carrying a mobile handset leg, measured 7 and 9 September 2026. Per-call results are provided in the pilot report.
  2. 2.False positive figures cover 12 genuine calls on clean lines and 40 recordings of genuine speech under degraded conditions, scored through the production pipeline under the same rule.
  3. 3.Trueline identifies synthetic speech. It does not authenticate a caller's identity and is intended to operate alongside existing authentication and fraud controls, not in place of them.