Hear it. Read it. Cite it.

Upload a recording or dictate straight into the browser and Akoé converts the audio to text automatically (automatic speech recognition or ASR, if you like the technical term): speech becomes written words, with speaker labels and a timestamp on every word. You then check the transcript in the editor, word by word, and export it as Word, subtitles, plain .txt file or TEI XML.

It depends on the recording: clear speech near the microphone needs lighter fixes; audio captured in noisy environments with bad use of microphones and in a low-resource language needs more. The next question lists what makes audio hard to transcribe. We do not guarantee an errorless transcript. What we promise is the ease of checking and correcting it: words for which our model returns a low confidence score get flagged for you to check, and the audio plays from the word you're checking at a speed you find comfortable, so correcting the transcript gets as easy as possible.

Recordings with distant or shared microphones, people talking over each other, switching languages mid-sentence, in noisy rooms and in languages, dialects or accents that are underrepresented in public training data are all factors that impact the quality of the resulting transcript. That last one is a dataset problem, not a defect of how anyone speaks. In the 'Languages' section above, you will find which languages are more likely to lead to better transcripts. We're based in Greece, a country rich in exactly the varieties that datasets underserve, so we feel this one personally—and we encourage research on lower-resourced speech. If your work needs a custom model for a specific variety or domain, write to us: describe your use case, and tell us whether you have recordings you hold the legal rights to share for training. Your normal uploads are never used for training—a custom model is a separate, explicit, written agreement.

No. Akoé transcribes speech, the words. It doesn't detect or mark non-speech sounds (laughter, coughing, crying), tone or emotion, and it doesn't transcribe silence: quiet stretches simply produce no text. If your method needs those cues, the editor is a good place to add them by hand—every word is timestamped, and the audio is right there while you work.

Yes, automatic speaker separation (speaker diarization) labels each voice, and telling the wizard how many people are speaking makes it sharper. Where it slips, the editor repairs it in a click: rename, reassign, split or merge speakers.

Transcription and dictation: about 100 languages, auto-detected or set by you—the languages section above shows what to expect from each. Speaker separation: any language. It works on voices, not words. Translation: into 28 languages, from a transcript in any of them: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish and Swedish (22 of the 24 official EU languages), plus Arabic, Catalan, Chinese, Norwegian, Russian and Turkish. Subtitles and exports: Word, subtitles (SRT and VTT), plain text, CSV and TEI, in the recording's language and in any translation you've made. Search by meaning: all transcription languages, and across them: ask in one language, find the moment in another. Mentions: people, places, organisations, dates and labels you define. Multilingual, indexed in the recording's language.

First your file waits for a processing machine—usually a couple of minutes. Then the transcription itself runs several times faster than the recording, so an hour of audio means minutes of processing, not another hour. On longer files the first minutes of text appear while the rest is still working, and we email you when the full transcript is ready—no need to keep the tab open. Sometimes a very short file doesn't feel proportionally fast, because the first wait is the same whether the recording lasts two minutes or two hours.

Yes—mp4, mov and other common formats. We keep only the audio track; the video is never kept. In modern browsers the sound is extracted on your device before upload, so a 2 GB interview video travels as a few megabytes of audio; if your browser can't do this, the file is converted on our EU servers and the original is deleted the moment the audio is out.

Audio: mp3, wav, m4a, flac, ogg/opus and most other common formats. Video: mp4, mov, webm etc.—we keep only the sound. If a file can't be read it isn't processed, and any charge refunds itself automatically.

Yes. Dictation runs straight in the browser: press record, speak, stop, and the text comes back fast — ready to play back word by word and export like any other recording. One voice, no speaker labels, the same €5-per-hour meter. It's also the fastest way to try Akoé: your first 5 minutes are free, no file and no card needed.

€5 per hour of audio, pro-rated to the second, with the exact price shown before you pay. The smallest checkout is €1 (card fees), and a €1 checkout covers up to 12 minutes of audio in the next 24 hours — your file included. Failed files refund automatically. Your first 5 minutes are free, no card needed.

Card and payment fees make charges under €1 not worth it for anyone. When the minimum exceeds your job's price (say a 5-minute memo) the difference covers more audio over the next 24 hours, your file included, so nothing is wasted. It lapses on its own and is never stored credit.

You could, the transcript is yours, export it any time and use it as you see fit. But the built-in translation is included with every transcript and keeps what a chat window loses: the draft stays aligned line by line with the original for easier review and every timestamp and speaker label survives.

Everything runs on EU-owned, EU-based infrastructure, publicly listed. Files are encrypted at rest. Retention is a schedule you choose and immediate deletion is always available. For your DPO or ethics committee, an Art. 28 data-processing agreement and the full process concept are available on request—and the security page carries the plain-language summary.

No. Your audio and transcripts are never used to train models, and no third-party AI service touches them. The models run self-hosted on EU servers.

Word (DOCX) in interview or dictation layout, plain text, SRT and VTT subtitles, CSV for spreadsheets, and TEI XML for archives — plus a CSV of the mentions index. Every export reflects your corrections, timings included.

Yes. The transcript is yours to keep and to publish. Researchers often add a full interview transcript as an appendix, and Akoé is made for it: export the corrected text as Word or plain text, with speaker labels and every word timed, so a reader can trace any quote back to its moment in the recording.

Yes, search finds moments by meaning—not just exact words, across your recordings. Your query can even be in a different language than the one of the recording. Results are real snippets from your own transcripts, each with a timestamp: click one and it plays. Your searches are never logged or stored, a transcript's search index is deleted together with it, and search is included with every account.

No. Akoé is pay-as-you-go automatic transcription and dictation app. You pay per recording, the price is shown first, and there is nothing to cancel: no money parked with us, no top-ups, no monthly fee.