AI transparency

How a transcript is produced

Transcripts are produced fully automatically: the uploaded recording is processed by a speech-recognition (ASR) model that converts sound into text, splits it into timestamped segments and approximately attributes utterances to speakers. No human listens to the recording or reviews the result — the text you see is the raw model output until you correct it yourself.

The cleaned-up version

Alongside the verbatim version you can ask for a cleaned-up one: the same text with fillers (“uh”, “um”, “you know”), repeated words and false starts removed. It is prepared by a language model, not by a person. It never starts on its own — it is produced only when you ask for it on a specific recording, because that is the one moment when transcript content is sent outside Cloudflare infrastructure.

The model is only ever allowed to delete. Before a cleaned-up version is stored, every passage is checked automatically: if the model reordered words, added anything of its own, deleted a longer stretch in one place than a single interjected phrase, or stripped more than half the words from an utterance, its answer is rejected and that utterance is left as it stands in the verbatim version. This is not a guarantee, however — the verbatim version remains the authoritative one, and the “Comparison” view shows plainly what was removed.

Models and providers

• ElevenLabs Scribe (ElevenLabs, Inc., USA) — the primary speech-recognition model. It receives the recording through a temporary, expiring link and returns the transcript; data transfers rely on Standard Contractual Clauses.
• Deepgram Nova-3 running on Cloudflare Workers AI infrastructure — the fallback speech-recognition model, used when the primary provider is unavailable. In that speech-recognition mode the recording does not leave Cloudflare infrastructure and is not passed to Deepgram, Inc. That statement covers speech recognition only — it does not extend to the cleaned-up version described below.
• Claude (Anthropic PBC, USA) — the language model that prepares the cleaned-up version of a transcript. It runs only when you ask for it on a specific recording; it then receives the transcript text (never the audio file). The provider retains that text as a rule for 30 days — longer only where its legal obligations or safety procedures require — and does not use it to train models. Transfers rely on Standard Contractual Clauses.

The current list of data processors and the rules for transfers outside the European Economic Area are set out in the Privacy Policy.

Limitations and accuracy

Model output contains errors. Accuracy depends on recording quality, the number of voices and how much they overlap, speech rate, dialect, specialist terminology and proper names. Speaker attribution is approximate and can confuse similar voices. We do not guarantee any particular accuracy level.

Verify a transcript against the recording before using it in official or court proceedings, in medical records, or for any other purpose where an error would have material consequences. Where the law requires a human-made transcript, our service does not replace one.

Your data and the models

We do not use your recordings or transcripts to train models, and we require the same of every model provider we use — both for speech recognition and for the language model that prepares the cleaned-up version. Data retention and deletion — including the automatic expiry of audio files and the retention period on the side of the cleaned-up version's provider — are described in the Privacy Policy.

How we mark AI content

• A visible notice next to every transcript: in the app, in the no-account preview and on the shared-transcript page.
• An informational sentence in the email notifying someone of a shared transcript.
• A footer line in the exported TXT file.
• Machine-readable marking: an aiGenerated field in the account data export (JSON) and metadata on the shared-transcript page.
• The cleaned-up version is marked separately and more emphatically: it is described as AI-generated everywhere it appears, the downloaded file's name tells it apart from the verbatim one, and its footer in the TXT file says plainly that a model altered the text rather than merely transcribing it.

The service is not a high-risk system within the meaning of the AI Act, performs no emotion recognition or biometric categorisation, and is not a conversational system. A transcript on its own reproduces the content of the recording without substantially altering the input data or its meaning, and we mark it for full transparency. The cleaned-up version, by contrast, is text a model altered, so we treat its marking as an obligation under art. 50(2) of the AI Act rather than as good practice.

What to keep in mind when using transcripts

• You are responsible for verifying a transcript before using it — especially proper names, numbers and utterances attributed to speakers.
• If you publish a transcript to inform the public on matters of public interest, you may have your own obligation to disclose that the text was artificially generated (art. 50(4) of the AI Act), unless the text has undergone editorial review and a natural or legal person holds editorial responsibility for its publication.
• You are responsible for the legal basis of recording and processing other people's statements — the Terms of Service and the Privacy Policy cover the details.

Contact and oversight

The service is provided by Wondel.ai sp. z o.o., registered in Warsaw, Poland, ul. Twarda 18, 00-105 Warszawa, KRS 0001029516, NIP (VAT) 5252951298, REGON 524989057. Questions about how AI works in the service: hello@wondel.ai.

Market surveillance of artificial-intelligence systems in Poland is carried out by the Commission for the Development and Safety of Artificial Intelligence (KRiBSI), established by the Polish act on artificial-intelligence systems implementing the AI Act. For personal data, the supervisory authority remains the President of the Personal Data Protection Office (UODO).