Questions that come up.

Real-time, two-way speech translation on a live phone call. Your agent speaks; the customer hears it in their own language in real time, on the same call leg, with no third human on the line. Both sides understand each other directly.

The pipeline (STT → translate → TTS) runs at the SIP boundary. See How It Works for the full anatomy of a call.

A human interpreter is one scheduled person, one call at a time — and you wait on availability. TransVoix is always-on across every call and language pair, in each speaker's own voice, at typically about half your contracted per-minute interpreter rate.

It's built for the full range of conversations — not a low-stakes carve-out. Teams that still want a human in the loop for a specific moment can keep one, but that's their choice, not a platform limit. See Pricing.

Audio is intercepted at your SIP boundary, processed in our US region, and returned to the call leg. No audio leaves the US. The audio path itself is stateless — frames are processed in memory and discarded.

The full subprocessor path (which vendors touch audio vs. metadata only) is documented on Compliance.

Inbound and outbound today. Multi-party / conference is on the roadmap and works in pilot deployments today with a single target language. If you need it, mention it during your scoping call.

Six languages live in production today (English, Spanish, French, Mandarin, Arabic, Hindi), for 30 bidirectional pairs. Eight more are next: Tagalog, Vietnamese, Korean, Portuguese, German, Russian, Italian, and Japanese. New languages ship on a published roadmap.

The full matrix — quality and speaker-count notes per language — lives on Languages.

An LLM-judge benchmark across the 30 production language pairs: average 4.72/5, scored 1 to 5. The methodology and the per-pair scores are on Languages.

Yes. Voice cloning, recording, and biometric capture all require fresh affirmative consent each session — never carried over from a stored preference. Consumer voice profiles are deleted on call end or consent revocation, whichever comes first; per-pilot agent voices are retained per the pilot agreement and are never reused across tenants.

The consent and provenance posture is detailed on Compliance.

Translation routes to dialect-aware variants where it matters (e.g. MX vs. ES Spanish, Brazilian vs. European Portuguese). Code-switching mid-utterance is detected at the phrase level and translated in place.

By default the audio path is stateless — frames are processed in memory and discarded, not written to disk. Recording is opt-in per tenant; when enabled, recordings are encrypted (AES-256) with per-call signed URLs that expire in 7 days.

The full retention matrix is on Compliance.

No. Customer call content is never used to train models. It's a hard line — verified by data-egress tests in CI, not just a policy doc. See the “Lines we don't cross” block on Compliance.

Sensitive tables (call recordings, consent audit, call turns) enforce service-role-only Row-Level Security, scoped to the originating tenant. The anonymous database key cannot read or write protected content — verified by a curl probe after every migration.

Consent and access events are written to an append-only audit trail — every grant, revoke, and expiry is a separate row, no UPDATE, no DELETE — so the state of access at any past moment is provable.

TLS 1.2+ on every endpoint, with WebSocket audio frames carried over WSS. AES-256 at rest; recordings encrypted with per-call signed URLs (7-day expiry). Provider API keys are stored as encrypted platform secrets — never committed to source, never logged.

Encryption and key-management posture is on the Compliance page.

We operate under HIPAA-aligned controls: encryption, access logging, breach notification, minimum-necessary data handling. For healthcare, we take your compliance team through the BAA and exactly what it covers; each subprocessor's BAA is executed before it receives any PHI.

We don't claim “HIPAA certified.” The full posture is on Compliance.

Our security program runs to SOC 2, our policies are in operation, and we've engaged an independent auditor for our SOC 2 Type 2 examination. We share the report under NDA once it is issued. We don't say “certified” until it is.

We can take you through where it stands under NDA, and the pen-test summary is available the same way: request either via the receipts list on Compliance.

For healthcare, we take your compliance team through the BAA and exactly what it covers. The DPA covers GDPR-scoped customers; custom redlines are accepted on Enterprise. Email security@transvoix.ai.

Calls are processed in the US today; for data outside the US, talk to us.

Call metadata (timestamps, duration, language pair, agent ID, tenant ID, error codes) for billing and debugging. Recording is opt-in per tenant. Voice-profile embeddings for any enrolled agents.

We do not keep transcripts, translation outputs, or user PII unless explicitly contracted into a retention plan.

Yes — per-tenant deletion controls, with the deletion event captured in the append-only audit trail. Full-tenant export and purge on contract termination is documented in the DPA.

Advance notice goes to your primary security contact before a new subprocessor that touches audio or PHI. The current named-vendor subprocessor list is available under NDA via the receipts list on Compliance.

On contract termination there's a contractual export window, then a purge across primary, replicas, and backups. The exact mechanics and timelines are in the standard DPA.

SIP-native architecture. Built for Amazon Connect, and designed to integrate with Genesys, Five9, NICE CXone, and other SIP-capable contact centers per customer engagement.

Provisioning timing scales with the agreed-upon scope and size of the pilot. We need SIP credentials, one disclosure script, and a list of pilot agents.

Scope, duration and success criteria are set with you on the scoping call. See Pricing.

A REST + WebSocket API for session control, agent enrollment, and audit-pack retrieval. The browser companion is the delivery surface today; a client SDK and native SIP are on our roadmap.

The audio segment on your agent's side of the call — distinct from the customer leg. Billing is metered on audio processed end-to-end on the agent leg only; idle time and silence over 500ms aren't billed.