Your old dialer already produces the data.Audrio makes it usable.
Every call transcribed, separated by speaker and searchable — in Polish and in English, detected per call. Audrio sits beside the phone system you already have: recordings arrive over an open API, nothing on the dialer changes, and it needs neither your CRM nor a project to install.
Open REST API and signed webhooks. Self-hosted or Audrio Cloud. EU and on-premise model endpoints. Independent of Salesforce. The live demo needs no signup.
Four steps, one deterministic pipeline
Ingest
Post a recording to /v1/calls or hand over a URL. Only call_id is required — every other field is optional, so nothing on the dialer has to change. Repeat uploads are idempotent.
Transcribe
Language is detected per call. Word-level timestamps, punctuation and sentence splitting, with speakers separated by channel mapping where the recording is split and by diarisation where it is not.
Analyse
A batch pass extracts topic, objections, pain points, purchase drivers, churn reason, department and funnel stage — every label carrying the quote and the timestamp it came from.
Search
Keyword, semantic and relationship search across every call, with results that deep-link to the exact second of the recording they came from.
Open a call and read it back
Everything here is synthetic — generated scripts and synthesised audio. No customer recording has ever entered this environment.
The third result is a Polish call answering an English question — one shared vector space across both languages.
Why deals were lost
Ranked, not scored. Counted across all 36 sample calls, from the resignation reason each lost deal was labelled with.
What it does
Transcription that survives real calls
Background noise, people talking over each other, compressed telephony audio, mono mixes. Custom vocabulary per language keeps house models, investment projects, place names, competitors and project codes intact.
Search that understands the question
Keyword search for names and codes, semantic search for situations — objections, complaints, financing worries — and relationship search across calls: every touchpoint for one customer, or which advisor knows a given project best.
Dashboards and compliance monitoring
Topic and objection trends, team and advisor comparison, funnel distribution, and a compliance view showing which scripted steps were followed, with review lists for quality staff to confirm or overturn.
An agent workbench
Ask in plain language — which calls mentioned loan worries last week — and get an answer with citations back to the recordings. Weekly and monthly reports generate themselves on a schedule.
Reliable before it is clever
A deterministic pipeline
Transcription, redaction, analysis and dashboard figures come from a scheduled batch pipeline, not from an agent improvising. Every result is auditable and can be re-run.
Overnight turnaround
Calls that arrive during the working day are transcribed, analysed and searchable by the next morning, so processing runs off-peak.
Retries you can see
Failures retry automatically and then surface in a dead-letter list an administrator can inspect and re-run from the interface.
Predictable cost
Per-call processing cost is budgeted and alerted on, and workbench usage is metered per user, so nothing runs away quietly.
Polish and English, both native
Polish is not a localisation of an English product here. Language is a property of the call, detected on arrival and carried through transcription, indexing, vocabulary and analysis. Adding another language is configuration, not architecture.
Detected per call
Language is identified on arrival and can be overridden by your metadata. Mixed calls take their dominant language.
Vocabulary per language
House models, project codes, districts and competitor names are maintained separately for Polish and English, with shared entries for codes.
One shared vector space
An English question finds the Polish call that answers it. Keyword search is tuned per language, including Polish diacritics.
Who uses it
Sales advisor
Reads back their own calls and reviews how objections were handled.
Team manager
Compares advisors and teams, follows objection and churn trends, and asks the workbench for representative calls.
Quality and compliance
Checks scripted steps, reviews flagged calls, and overrides the machine verdict where needed.
Operations and IT
Maintains vocabulary and topic packs, roles and audit, watches pipeline health, and configures model endpoints.
One endpoint to integrate against
Your IT team builds the dialer integration against a documented OpenAPI specification, with example scripts. Completion is announced by an HMAC-signed webhook; a scheduled SFTP or shared-folder pull is available where pushing is not an option.
Ask for the API referencecurl -X POST https://api.audrio.io/v1/calls \ -H "Authorization: Bearer $AUDRIO_KEY" \ -H "X-Audrio-Signature: $HMAC" \ -F "call_id=demo-en-003" \ -F "audio=@2026-07-24-093112.wav"
{ "call_id": "demo-en-003", "status": "received" }Self-hosted or Audrio Cloud
| Self-hosted or Audrio Cloud | Self-hosted | Audrio Cloud |
|---|---|---|
| Data repository | Your environment, under your control — audio, transcripts, indexes and graph | Managed by Audrio. An EU-resident Cloud is on the roadmap |
| AI processing | Every endpoint is configurable — transcription, inference, embedding and reranking each point at an EU-region hosted endpoint or at your own on-premise models, in which case nothing leaves your network | Endpoints configurable per instance. An EU preset is on the roadmap |
| Deployment | A single Docker Compose delivery, upgraded file by file | Nothing to run |
| Best for | Organisations with strict GDPR requirements | Teams who want to start without infrastructure |
Two separate questions. Where the data rests: self-hosted today, with an EU-resident Cloud on the roadmap. Where the models run: every endpoint is configurable, so both routes work today — EU-region hosted endpoints where processing stays inside the EEA, or your own on-premise models where nothing leaves your network at all.
Privacy enforced in code, not only in policy
Redaction before analysis
Phone numbers, e-mail, PESEL, NIP, IBAN, names and addresses are removed before any text reaches a model. Raw text needs a second authorisation and is logged.
Endpoints you choose
Transcription, inference, embedding and reranking each point wherever you tell them to — an EU-region hosted endpoint, or your own on-premise models, in which case nothing leaves your network.
Signed and gated
API keys, HMAC request signatures, HTTPS only, CORS allow-lists, and an authentication gateway in front of the workbench. Per-key request limits keep a runaway script or a stolen key from flooding the pipeline.
Read-only for the agent
The workbench reaches the intelligence layer through a read-only tool allow-list. It can answer questions; it cannot write.
Deletion on request
Locate and erase everything tied to a call or a phone number, including derived indexes and graph edges.
Everything logged
Viewing raw text, exporting, editing vocabulary and changing configuration are all recorded in the audit log.
Test it on your own recordings
Send us a set of your real Polish calls — background noise, interruptions, mono mixes and all. We run them through the pipeline and hand back the transcripts, the speaker separation, the extracted objections and the search results, so you can read the output line by line before committing to anything.
Asked before
Does Audrio replace our phone system?
No. Audrio sits beside it. Recordings arrive over the API and nothing on the dialer changes.
Do we need Salesforce?
No. Audrio analyses calls on its own. CRM integration is a later phase, not a prerequisite.
Can you handle mono recordings?
Yes. Where the advisor and the customer sit on separate channels, separating them is direct. Where they are mixed into one track, Audrio identifies from the audio where the speaker changes and which voice is the advisor. Both are supported — you do not need to check your recording format first.
Where does our data live?
The repository is self-hosted in your own environment today; an EU-resident Cloud is on the roadmap. AI processing is a separate setting: point each endpoint at an EU-region hosted service, or at your own on-premise models — in which case nothing leaves your network at all.
Can we add our own vocabulary?
Yes — per language, in the interface or by CSV import: house models, project codes, districts, competitors.
How soon do we see real output?
The proof of concept runs on your own recordings first. Nothing is committed before you have read the results.