Frequently asked questions

Short answers to the questions teams ask before and after connecting VoiceLens: agents and bots, languages, accuracy, integrations, pricing and data.

Product

Does VoiceLens work with human agents, AI agents, or both?

Both, with the same rubric. Detectors look at speech roles (agent turn, customer turn), not at who produced them. Tag each call with agentType: "ai" or "human" in the ingest metadata to compare the two populations in analytics.

Which call platforms are supported?

Any platform that can give you a recording. The ingest API accepts a file, an https URL or a pre-signed upload, and Twilio recordings can be sent automatically through a status callback. Plivo, Vonage and custom providers are supported through the generic API.

Which languages are supported?

Transcription covers nearly 60 languages and follows the language set on the upload, the language set on the channel, or automatic detection, in that order. Detection itself only recognises about 35 of them and labels anything else as the nearest language it knows — Slovenian usually arrives as Bulgarian — so declare the language when you know it. Fault detection lexicons exist for English, German, Spanish, Italian, Arabic and Turkish; calls in other languages are transcribed in their own language but graded with the English word-based detectors. Sentiment scoring is currently English only. The dashboard and this documentation are available in those six languages.

How long does analysis take?

Usually well under a minute for a typical call; long recordings take proportionally longer. Subscribe to the conversation.completed webhook rather than polling.

Can I analyse transcripts without audio?

Yes. The dashboard accepts raw Deepgram JSON output. Everything works except the acoustic line-quality measurement and audio playback.

Accuracy

How accurate is the health score?

It is as accurate as the rubric behind it. Build a golden set of hand-scored calls and run calibration; the report gives you mean error, bias and the share of calls within ten points. Most teams reach a mean absolute error under ten points after two or three rubric iterations.

What if a fault is wrong?

Mark it as a false positive on the call. It is excluded from that call's health score, and the fault precision dashboard tracks the false-positive rate per code so you can see which detectors need adjusting or disabling for your call type.

Does a bad phone line hurt the agent's score?

No. Line quality, low transcription confidence and speaker uncertainty are info-level faults. They are shown so you know how much to trust the transcript, but they never lower the health score.

Integration

How do I get results into my own tools?

Three ways: webhooks push each completed analysis to your URL, signed with HMAC-SHA256; a read-only API key lets your reporting pull conversations and analytics; an embedded dashboard shows the data inside your portal.

Are ingest tokens dangerous if leaked?

They can only create calls in one channel; they cannot read anything. Revoke and rotate anyway, and set an IP allowlist for fixed infrastructure.

Do you deduplicate re-sent recordings?

Yes, on Idempotency-Key, on externalCallId per channel, and on Twilio's RecordingSid. Duplicates return the original conversation and do not consume minutes.

Pricing

How are minutes counted?

By the full duration of each analysed call, whether it arrived as audio or transcript. Skipped and failed calls are not metered. Re-evaluating a call with a new rubric is free; re-transcribing is not.

What happens when I run out of minutes?

The upload that crosses the limit completes; the next one is refused with 402 usage_limit_reached until the period resets, you upgrade, or minutes are added. Pro and Agency accounts can add minutes at $0.05 and $0.04 per minute respectively. See Usage and billing.