Ses Notu
An iOS app that transcribes in-person meetings from the middle of the table, separates who said what, and turns the conversation into a summary with decisions and action items.
Screenshots in Turkish and English, app icon. Everything on this page is free to use in coverage.
Quick facts
| Developer | Utku Bayık — independent, one person, no outside investment |
| Platform | iPhone and Apple Watch (iOS). No iPad, Mac or Android build. |
| Languages | Turkish and English |
| Price | Free for the first week, then Ses Notu Pro monthly subscription |
| Category | Productivity |
| Funding | Four Turkish public grant programmes: two TÜBİTAK, two KOSGEB |
| App Store | apps.apple.com/app/id6802013702 |
| Website | sesnotu.web.app |
Boilerplate copy
Use any of these as-is. No approval needed.
English
Ses Notu transcribes in-person meetings from the middle of the table, separating who said what.
Ses Notu is an iOS app for meetings where everyone is in the same room. Put your phone on the table and the conversation becomes text live, with each speaker separated as they talk. When it ends you get a summary, the decisions, and action items with owners and dates.
Almost every meeting assistant on the market works by sending a bot into Zoom, Meet or Teams, which does nothing for the meeting where everyone is already sitting around the same table. Ses Notu is built for that meeting. You put your iPhone in the middle of the table and tap record; speech becomes text as people talk, and speakers are separated from each other — Speaker 1, Speaker 2, Speaker 3 — with names mentioned during the conversation matched to the right voice. Afterwards you get a summary, the decisions that were made, and action items with an owner and a due date; "by Friday" is resolved into a real date and can go to Reminders with one tap. You can ask the transcript questions later and answers cite the moment they came from.
Türkçe
Ses Notu, masadaki yüz yüze toplantıyı yazıya döker ve kim ne dediyse ayırır.
Ses Notu, herkesin aynı odada olduğu toplantılar için bir iOS uygulaması. Telefonu masanın ortasına koyuyorsunuz, konuşma canlı olarak yazıya dökülüyor ve konuşmacılar birbirinden ayrılıyor. Toplantı bitince özet, alınan kararlar ve sorumlusuyla terminiyle birlikte yapılacaklar listesi hazır oluyor.
Piyasadaki toplantı asistanlarının neredeyse tamamı Zoom, Meet ve Teams'e bot sokarak çalışıyor; herkesin aynı masada oturduğu toplantı çözümsüz kalıyor. Ses Notu tam o toplantı için yazıldı. iPhone'u masanın ortasına koyup kayda basıyorsunuz; konuşma insanlar konuştukça yazıya dönüyor ve konuşmacılar birbirinden ayrılıyor — Konuşmacı 1, Konuşmacı 2, Konuşmacı 3 — toplantıda geçen adlar doğru sese eşleniyor. Sonunda özet, alınan kararlar ve sorumlusuyla terminiyle yapılacaklar çıkıyor; "cuma gününe kadar" gerçek bir tarihe çevriliyor ve tek dokunuşla Hatırlatıcılar'a gidiyor. Sonradan transkripte soru sorabiliyorsunuz, cevaplar geldikleri anı kaynak gösteriyor.
What's actually different
- It is not a meeting bot. There is no call to join. The phone sits on the table between the coffee cups.
- Speaker separation runs on the device. The voice embeddings used to tell people apart never leave the phone.
- Turkish works properly. Otter, the category leader, does not support it. Otter's own help centre, updated May 2026, lists six languages: English, Spanish, French, German, Japanese and Chinese (Simplified). This gap was first written about in Turkish tech press in 2020 and is unchanged six years later.
- It survives real conditions. Recording continues with the screen locked and resumes after an incoming phone call.
- Apple Watch and Live Activities. Start and stop from the wrist; the recording stays visible on the Lock Screen and in the Dynamic Island.
Technical notes
For writers who want specifics rather than marketing language.
- Speaker embeddings come from CAM++ (WeSpeaker), trained on VoxCeleb, running locally through ONNX Runtime. Apache-2.0 licensed. About 28 MB, shipped inside the app.
- Apple provides no speaker recognition API — the
Speechframework has nothing for it — which is why the model is bundled rather than called. - The model takes 80-band Kaldi-compatible filterbank features, not raw audio: 25 ms window, 10 ms hop, 16 kHz, followed by cepstral mean normalisation. Output is a 192-dimension vector, L2 normalised, compared by cosine similarity.
- Measured against real recordings: two sentences from the same person score 0.83 similarity; different people score 0.20.
- ONNX Runtime was chosen over CoreML deliberately. The work happens after recording on clips of a few seconds, not in real time, so the CPU is sufficient and the conversion cost was not worth the Neural Engine.
Privacy, stated accurately
This is worth getting right rather than overclaiming, so here is the full picture.
- Stays on the device: recordings, transcripts, and the voice embeddings used to identify speakers. Voiceprints are biometric data and are never uploaded; they can be deleted from inside the app.
- Leaves the device: audio is streamed to Soniox to produce the transcript, and transcript text goes to Google Gemini for summaries and the chat feature.
- Recordings are not uploaded to any account. If the subscription lapses, everything already recorded stays fully accessible.
- Telling the people in the room that they are being recorded is the user's responsibility, and the app says so.
Screenshots
Full resolution, 1320 × 2868. Click to open, or take everything in the zip.
Video
Roughly 25 seconds, 886 × 1920, no narration. English · Turkish
Press contact
Utku Bayık, developer. Happy to answer technical questions, provide review codes, record a specific flow on video, or talk through how the on-device diarization was built.