Stress the real listening conditions.
Multiple voices, accents, genders and speaking rates, with optional trading-floor background noise, challenge recording and transcription beyond ideal audio.
PRODUCT / ASSURE
Evidence follows
the call.
Recording-path assurance built on greenstar’s automated call-testing expertise.
Download Assure brief · PDF ↓INSIDE ASSURE / THE VOICE TEST ENGINE
The right words.
On the right leg.
Not just a file that exists. Proof of what the conversation captured.
Assure generates and verifies real, two-way spoken test conversations. Text-to-speech creates the caller. Speech-to-text listens at both ends and checks the recovered recording. Answer classification establishes who—or what—picked up.
Built on greenstar’s TTS, STT and answer-detection engine, these checks establish that the right words were captured, on the right leg, at usable quality, transcribed accurately and flagged by surveillance.
Meet the caller ↓ Meet the verifier ↓ Check who answered ↓ENGINE / 01 / TEXT-TO-SPEECH
A synthetic caller.
A recognisable conversation.
Every test call carries a unique spoken token.
The engine speaks the line, UTC time and call token into the conversation. That identity stays in the audio and transcript even when call metadata is lost. Both ends speak distinct phrases, so each leg can be proven independently.
Two-sided scripts use markets vocabulary—contract months, “yard”, bid/offer and “done”—to test the language a desk actually uses, not just a clean tone.
Hear the identity in the words.
- Caller / leg A
- “TraderVoice test, line 4417, 14:32:05 UTC, token bravo-seven-niner”
- Responder / leg B
- “TraderVoice responder, token delta-four-two. September contract, confirm bid/offer.”
Illustrative script, not a live call or a measured result.
Say what the control should catch.
Authorized surveillance scripts deliberately include phrases such as “let’s take this to WhatsApp” or a promise of returns. Assure checks that the expected alert actually fired—not merely that the phrase was recorded.
ENGINE / 02 / SPEECH-TO-TEXT
Listen live.
Verify the recovered record.
Did each end hear the other?
Each end transcribes the other in real time. A missing far-side token immediately flags one-way or no audio. A connected call alone is not a successful audio check.
Did the whole script survive?
Assure retrieves the recording from the platform or archive and transcribes it. Expected tokens and script must be present on the correct legs, in order and complete, without clipping or truncation.
How accurate is the transcript?
The known script is ground truth. Assure measures the firm’s own archive or surveillance transcription against it using word error rate (WER), making transcription quality a testable control.
Quality in context.
STT confidence and word error rate on recovered audio sit alongside mean opinion score (MOS), jitter and packet loss. Together they show whether speech was usable, not simply whether a recording exists.
Timing from mouth to ear.
The time between a token being spoken and heard measures mouth-to-ear latency per leg. On hoots, Assure measures one speaker against every subscriber, checking fan-out and per-endpoint delay rather than assuming all listeners heard the same thing.
ENGINE / 03 / MACHINE VS HUMAN
An answer is not
the expected answer.
Only the expected answering party can unlock a verified result.
Every answered call is classified: live human, voicemail, IVR / auto-attendant, special information tone (SIT), fax / modem tone, dead air or the TraderVoice test responder.
If a human trader answers, Assure plays a short courtesy message, ends the call, logs the event and reschedules. It protects the desk from an unwanted test conversation; reaching a human is not counted as reaching the test responder.
The expected answer is a gate, not the whole test. Recording content, quality and surveillance checks retain their individual outcomes, so an answered call cannot hide a missing return leg or a missing alert.
Identify the test. Keep the evidence.
Tokens and answer classification identify test calls so firms can auto-tag them and exclude them from human surveillance review while still proving capture and checking that the expected surveillance alert fired.
A foundation for deepfake defense.
The same machine-vs-human models underpin scoring of live inbound calls for synthetic or cloned voices. The labeled TTS library also provides training and red-team material. The shared engine connects assurance and fraud defense; a voice-risk score is a signal for review, not proof of identity or authority.
Explore voice-risk evaluation →Assure is the first focus. Record, Desk and Agents remain on the product roadmap. Voice-risk evaluation considers false positives, false negatives and changing attack conditions alongside human verification.
01 / TEST DESIGN
Follow the whole recording path.
Test the path
Place authorized, scheduled calls with known audio across the voice routes in scope. Scope ringdowns, hoots, PSTN and mobile in your testing roadmap.
Verify the record
Retrieve the expected recording, compare the audio and inspect the configured recorder-to-archive path. Record failures and uncertainty explicitly.
Keep the evidence
Bring timestamps, source IDs, individual check results and exceptions into a reviewable evidence pack. A pass applies to the tested scope—not every call.
02 / CONTROL POSTURE
See the gap. Keep the context.
TEST POPULATION05 example paths
02 exceptions to review
| Line | Target platform | Call | Audio | Archive |
|---|---|---|---|---|
| LDN · 001 | Cloud9 | ✓ Pass | ✓ Pass | ✓ Pass |
| LDN · 002 | IPC | ✓ Pass | ! Gap | ! Pending |
| NYC · 003 | BT | ✓ Pass | ✓ Pass | ✓ Pass |
| NYC · 004 | Teams | ✓ Pass | ✓ Pass | ! Gap |
| CHI · 005 | SIP/PSTN | ✓ Pass | ✓ Pass | ✓ Pass |
! DEMO EXCEPTION / NYC · 004
Archive retrieval not confirmed. Review the ingest path before closing the exception.
03 / PLANNED CHECKS
More than a dial tone.
Known synthetic audio
Define scheduled calls and fingerprinted audio around the recording paths, permissions and testing windows your firm needs.
Recording and archive
Trace the expected record across capture and storage: both-leg audio, metadata, transcript linkage and retrieval. Distinguish access issues from missing recordings.
Line health and exceptions
Build a line-health workflow that separates failures from inconclusive checks and gives each exception a clear owner.
Retention context
A policy check can report the configuration it saw. It cannot promise a record will survive its required lifetime. Legal holds, retention schedules and retrieval need separate review.
Signed evidence packs
Define evidence integrity through source identifiers, reliable timestamps and a documented signing and key-custody process.
Voice-risk evaluation
Assess synthetic-voice risk through detection evaluation and labeled red-team exercises tailored to your channels.
04 / EVIDENCE MODEL
The limits belong in the report.
No universal green checkmark.
A useful pack makes it possible to reconstruct what happened. Test scope, original records, unknowns and the path to resolution belong together.
See the result, the exception and the next action in one place.
Read the regulatory boundaries →CONTROL TEST
TV-DEMO-0042
TV-DEMO-0042
One test. A traceable record.
- Test scope
- Demo desk / line 004 / test path A
- Sample timestamp
- 2026-09-26 14:32:08 UTC
- Source IDs
- CALL-DEMO-42 → REC-DEMO-42
- Call pathPASS
- Expected audioPASS
- Both-leg samplePASS
- Archive retrievalGAP
- Retention policyNOT TESTED
EXAMPLE SHA-256
c3dc20d5615b858a427eedafcb1029fd350442ae3bd89e54e008cd07a2239d48Example exception: archive record unavailable. Owner: reviewer. Next step: inspect the ingest path.
QUESTIONS / ANSWERS
Know what a test can—and cannot—prove.
Does Assure replace my recorder?
No. Assure is designed to test recording paths alongside existing systems. It is not a replacement for your recorder, archive or supervisory controls.
Do you access our recordings?
Recording checks use agreed, authorized access. Define permissions, deployment and data handling for your environment before testing.
Which platforms are supported?
greenstar technology owns existing Cloud9 automation for line and recording checks. The integration roadmap extends to IPC, BT, Teams, Zoom Phone, SIP/PSTN, mobile capture and archive systems.
How often do tests run?
Set a testing cadence around your firm’s risk assessment, recording architecture and authorized test windows.
What would an examiner receive?
An evidence pack brings together the tested path, UTC timestamps, record identifiers, check results and exceptions for review.
Does a passing test prove every call was recorded?
No. A synthetic test establishes a result only for that test, path, configuration and time. It cannot prove that every production call between tests was recorded.
Is fingerprint matching the same as deepfake detection?
No. Matching known test audio asks whether expected content reached a recording. Detecting unknown synthetic speech is a separate task; neither proves caller identity or authority.
How can I request a conversation?
Use the demo or early-access form to request a conversation. Online delivery may be unavailable during maintenance; you can also email info@tradervoice.app.
EXPLORE THE WORKFLOW
From the first test to the next action.
EXPERIENCE BEHIND ASSURE
Guided checks of the mobile experience.
greenstar technology has also built self-service Teams Phone Mobile test automation. Users initiate a call, receive an automated callback, and confirm ringing and recording announcements by voice or keypad. Refined through real-call feedback, the workflow extends our experience from infrastructure-level call testing to guided checks of the mobile user experience.
The mobile workflow captures user confirmation of ringing and recording announcements—not proof of recording capture or a released Assure connector.
NEXT / DEFINE THE TEST
Start with the paths
you need to trust.
Bring your recording architecture. Let’s define the right test.