Most organizations that record conversations end up sitting on an archive nobody can use. The calls are kept because compliance says so, or because someone might need them one day, and then they are never opened again. The value in them is real and completely locked up, since the answer to almost any question about what a customer actually said is somewhere in there, and finding it means someone scrubbing through hours of audio.
I built the Private Speech QA Agent to turn that archive into something a team can simply ask. It transcribes recordings into speaker-attributed, timestamped text, makes every conversation searchable, and puts a conversational agent on top, and it does all of that inside the organization’s own environment, which for most of the people who need this is the only version that is usable at all.
Asking questions across the whole archive
The agent takes ordinary questions in ordinary language, answered across every recording at once: which objections came up most often on last month’s calls, whether anyone promised a particular customer a refund and when, where the engineer explained the migration plan and what the summary of it is.
Follow-ups matter more than they look. The agent keeps the thread of a conversation, so a question like “what did they say right after that” resolves against the previous answer rather than starting a fresh search, which makes it behave less like a search box and more like asking a colleague who has actually listened to everything.
Why the recordings cannot leave the building
For the organizations that need this most, uploading the audio is the one thing they cannot do. Call centers, telecoms, customer support teams, and anyone under regulatory supervision are holding recordings full of personal data, commercial terms, and things customers said in confidence, so sending that to a third-party API is often a straightforward breach of policy, of contract, or of law. The archive stays shut and the value stays locked up.
This runs entirely inside the client’s own account. The recordings, the transcripts, the search index, and the model all sit in infrastructure they control, and nothing is sent to an external provider, which turns the compliance conversation into one about their own environment rather than about somebody else’s terms of service. That single property is what makes the rest of it possible.
Grounding every answer in the source audio
For a system answering questions about someone’s private recordings, a confident wrong answer is worse than no answer at all, so every answer is grounded in the actual transcripts and cites where it came from, down to the file and the moment in the recording. A user can click straight to the timestamp and hear it for themselves. The agent works only from what it retrieves, and when the recordings genuinely do not answer the question it says so rather than filling the gap.
That verifiability is the difference between a demo that impresses in a meeting and something a team will put real decisions on top of.
Who this is for
Any organization with a large and growing pile of recorded conversations that cannot leave the building. In practice that has meant telecommunications, media, customer support, and settings where regulatory compliance drives both the recording and the restrictions on it. The common shape is the same everywhere, in that recordings are already being captured for one reason, and there is far more value in them than anyone is currently able to extract.
Running on AWS in the client’s own account
The system deploys to AWS as infrastructure-as-code, so it arrives as a unit rather than as a bespoke installation. It can run fully self-hosted on open models, or be wired to managed models where an organization prefers that tradeoff, and the choice is a dial rather than a rebuild.
Project status
This was built as a proof of innovative concept. Over a twelve-month project running from March 2025 to March 2026, we developed the prototype, deployed it to AWS, and tested it end to end, demonstrating the technical feasibility of private speech recognition combined with retrieval-based question answering. Alongside the build we produced a market analysis and a commercialization plan, and filed a trademark application with EUIPO.
Stated honestly, the concept is proven in a controlled environment and the foundations are in place for development toward production readiness. It is not a shipped product with a customer list.
The full completion report, including project activities and financing, is published at privatesynapse.ai.
Funding
The Private Speech Q&A project (reference NPOO.C3.2.R3-I1.06.0015) was carried out by Intelligentia d.o.o. and co-financed by the National Recovery and Resilience Plan under the “Proof of Innovative Concept” call. Total project value 79,732.67 EUR, of which 49,555.36 EUR was EU co-financing.
Funded by the European Union, NextGenerationEU. The views and opinions expressed are solely those of the author and do not necessarily reflect the official positions of the European Union or the European Commission. Neither the European Union nor the European Commission can be held responsible for them.
Related: PrivASR, the private speech-to-text engine this is built on.