Today we are introducing Open Voice Shield, AI-powered fraud analysis for SIP traffic. Point a trunk at it and every call comes back with a verdict of block, review or allow, a fraud probability, the category, the red flags that earned the score, a stereo recording and a speaker-separated transcript. Live audio is never touched, nothing is installed, and every new account starts with US$5 of credit and no card.
Open Voice Shield is built and operated by TelecomsXChange. It lives at ovs.telecomsxchange.com and has its own page on this site at /open-voice-shield.
Why voice needed this
We have shipped SMS fraud detection for years with OpenTextShield, and the question we kept getting from voice operators was some version of: can you do that for calls? The honest answer used to be no, because the hard part of voice fraud is not the signalling. It is what was said.
A bank impersonation call and a legitimate support call look identical in a CDR. Same A-number format, same duration range, same codec. The difference is a caller telling someone their account will be frozen in an hour unless they read out a card number. You cannot see that from the SIP layer, and no operator has the staff to listen to calls at scale.
Two things changed. Transcription got good enough to separate speakers reliably on 8 kHz G.711 audio, and language models got good enough to read a transcript and answer a narrow question with a structured, repeatable output. Open Voice Shield is what you get when you put both behind a standard SIP proxy.
What a verdict contains
Every completed call produces a strict JSON verdict, which is the thing you automate on. It carries:
- Recommendation: block, review or allow.
- Probability: a score from 0 to 100.
- Category: the fraud type from a fixed grouping, plus the entity and caller identity claimed on the call.
- Red flags: urgency, fear tactics, PII harvesting and the other patterns that pushed the score.
- TCPA note: whether the call looks like the kind that carries $500 to $1,500 of per-call exposure.
- Title, summary, keywords and language. Calls in English, Spanish, French and more are analysed in their own language.
Three AI depths are available and you switch them in the portal: Economy for fast, lean models on every call, Power as the default daily driver, and Max for the most capable models. The choice applies to new calls instantly.
Nothing touches the live call
We want to be plain about the architecture, because it is the reason operators can trust it in front of production traffic. Open Voice Shield is in the media path. RTP goes through its media server in both directions, and that is how the call gets recorded. What it never does is inject anything into the audio or make the live call wait on analysis.
While the call is live, five things happen: the INVITE arrives at the OVS ingress, the source IP is checked against your CIDR whitelist along with account status, concurrency and balance, the SDP connection info is rewritten so media is anchored and a pcap is written per stream, the call is routed to your destination by longest-match prefix rule or account default, and a BYE from either side tears everything down and reports the metrics.
Only after hangup does the asynchronous pipeline start: locate the captures, decode G.711 to a stereo 8 kHz WAV, transcribe each channel with caller and callee labels, run the language model, then charge the fee, fire alerts and automation rules and post signed webhooks. If any step in that pipeline fails, the only thing that happens is a missing verdict. The call itself already completed normally.
Acting on it without a human in the loop
A verdict you have to read is a report. A verdict that acts is a control. Open Voice Shield ships with both.
- Automation rules block or divert a caller when the probability crosses your threshold, for a duration you choose. Blocked callers receive a SIP 403.
- Email alerts match on probability, category, recommendation or call quality, so a MOS ceiling can page you the same way a fraud score does.
- Webhooks are HMAC-SHA256 signed POSTs, so you can drop verdicts into your own switch, CRM or SIEM and verify they came from OVS.
- Fraud cases carry severity, assignment and a comment thread, and tickets can be opened straight from the call detail.
Evidence you can hand to someone else
Fraud on a voice route is expensive twice: once when it happens and again when you have to prove it to a carrier, a customer or a regulator. So every call keeps its evidence: the stereo WAV with the caller left and the callee right, the transcript, a timestamped SIP ladder with full header expansion, the raw media pcap (zipped for multi-leg calls) and estimated MOS, jitter and packet loss per leg with the worst leg shown on the CDR.
An analyst bundle packages all of that into a single zip with an HTML report, and read-only share links can be generated with a password and an expiry of 24 hours, 7 days, 30 days or never, with view counts tracked. Retention windows for recordings, pcaps and SIP logs are set independently, and storage can be offloaded to an S3-compatible bucket you own.
The wholesale angle: catching the vendor, not just the caller
If you buy termination, the fraud you lose the most money to may not come from the A-number at all. Delayed-playback false answer supervision is a vendor swapping the callee to a looping recording mid-call so the billing keeps running. It is nearly invisible in aggregate stats because the call did connect and did last.
Open Voice Shield watches the callee channel for exactly that switch from interactive conversation to playback and raises a route alert with a confidence score. Combined with per-leg MOS on every CDR, a degrading route shows up in the table before it shows up in a complaint. For buyers on the TCXC Marketplace, putting OVS in front of purchased routes turns a vendor dispute from an argument into an attachment.
Integration is three settings
There is no agent, no digest credentials and nothing to change on the caller side:
- Whitelist your IP. Add your SIP server public IP as a CIDR entry. Authentication is by source IP, per tenant.
- Add the destination. The host, port and transport of the switch or carrier OVS should forward to. Prefix rules are optional if different destinations need different handling.
- Point a trunk at the ingress. UDP or TCP on port 5060, shown on the Network page. Make a test call and read the verdict the same afternoon.
It works with any SIP-capable switch, SBC or PBX. Media is G.711 today; other codecs pass through but are not analysed yet. Custom SIP headers can be captured to indexed CDR columns, and a beta option forwards calls to voice AI agents.
Pricing
Open Voice Shield is prepaid. You pay per minute for the call and one fixed fee per analysed call, because an analysis is one piece of work regardless of how long the call ran.
| Plan | Monthly | Per minute | Included minutes | Concurrent ports |
|---|---|---|---|---|
| Pay-as-you-go | $0 | $0.02 | None | Unlimited |
| Starter | $249 | $0.018 | 15,000 | 24 |
| Business | $699 | $0.016 | 50,000 | 96 |
| Enterprise | $2,499 | $0.014 | 200,000 | 384 |
The analysis fee depends on the AI mode: $0.02 on Economy, $0.05 on Power and $0.15 on Max. A 3-minute pay-as-you-go call therefore costs $0.08, $0.11 or $0.21. Pay-as-you-go bills in 60-second then 1-second increments and has no concurrency limit; bundles bill in 6-second then 1-second increments, and leaving one mid-cycle credits the unused part. Full details are on the pricing page.
Start with one test call
Every new account starts with US$5 of credit and no card. Whitelist an IP, point a trunk, make a call, and see what your traffic looks like with a verdict attached.
Create your Open Voice Shield account →
Or read the product overview, the documentation, or talk to us first.
